paper_id
string
title
string
domain_dataset_origin
string
published_date
string
arxiv_categories
list
legal_license
string
license_url
string
commercial_use_allowed
bool
commercial_ip_safety_score
int64
commercial_ip_verdict
string
commercial_ip_justification
string
academic_citations_count
int64
influential_citations_count
int64
field_weighted_citation_impact_fwci
float64
tldr_neural_summary
string
primary_author
string
total_authors_count
int64
authors
list
affiliations
list
arxiv_url
string
code_audit_status
string
primary_repo_url
string
repo_urls
list
github_stars
int64
github_forks
int64
github_last_commit
string
github_license
string
importance_score
float64
abstract
string
title_vector_384d
list
abstract_vector_384d
list
target_industry
string
reproduction_recipe
string
core_problem_addressed
string
key_technical_innovation
string
key_quantitative_result
string
cluster_id
int64
cluster_topic_name
string
trend_velocity_tier
string
papers_last_30_days
int64
audited_at
string
2606.20717v1
MIRAGE: Stealthy Visual Prompt Injection for Vulnerability Detection in Web Agents
AI Security, Red Teaming & Defense 2026
2026-06-16
[ "cs.CV", "cs.AI", "cs.CR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Multimodal Large Language Model (MLLM)-based web agents provide practical, high-precision solutions for visual browser automation; however, they inherently expand the attack surface, introducing novel vision-based vulnerabilities. Existing adversaria
Xuelong Dai
6
[ "Xuelong Dai", "Jianyu Ma", "Boyang Ma", "Biwei Yan", "Yijun Yang", "Yue Zhang" ]
[]
http://arxiv.org/abs/2606.20717v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/browser-use/browser-use", "https://github.com/openclaw/openclaw" ]
386,158
81,165
2026-08-13
NOASSERTION
186.77
Multimodal Large Language Model (MLLM)-based web agents provide practical, high-precision solutions for visual browser automation; however, they inherently expand the attack surface, introducing novel vision-based vulnerabilities. Existing adversarial evaluations targeting these agents frequently rely on permissive thr...
[ 0.02351200021803379, 0.01979300007224083, -0.04714199900627136, 0.02040099911391735, 0.05547900125384331, -0.01540399994701147, 0.05035800114274025, -0.02896299958229065, 0.0019600000232458115, -0.047237999737262726, 0.03424699977040291, -0.10742399841547012, 0.10490699857473373, -0.020897...
[ -0.009228000417351723, -0.051242999732494354, -0.04223199933767319, 0.011796999722719193, 0.11453799903392792, -0.020932000130414963, 0.006814000196754932, -0.08849000185728073, -0.02828799933195114, -0.0928340032696724, 0.046358998864889145, -0.06631399691104889, 0.09932400286197662, 0.04...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Multimodal Large Language Model (MLLM)-based web agents provide practical, high-precision solutions for visual browser automation; however, they inherently expand the attack surface, introducing novel vision-based vulnerabilities.
Operating under these realistic constraints, we propose MIRAGE, a novel visual indirect prompt injection framework for targeted next-action hijacking.
Comprehensive evaluations against prominent MLLM web agent frameworks, specifically SeeAct and OpenClaw, empirically demonstrate the potency, realism, and stealth of our proposed MIRAGE.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:25:31.505330
2605.27117v1
Position: AI Safety Requires Effective Controllability
AI Security, Red Teaming & Defense 2026
2026-05-26
[ "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
AI safety is still largely framed as alignment: training models to follow human preferences, safety policies, and normative constraints. That framing has improved the behavior of modern language models, but aligned behavior does not by itself guarant
Yige Li
3
[ "Yige Li", "Yunhao Feng", "Jun Sun" ]
[]
http://arxiv.org/abs/2605.27117v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/openclaw/clawhub", "https://github.com/ECNU-ICALK/AutoSkill", "https://github.com/openclaw/openclaw" ]
386,158
81,165
2026-08-13
NOASSERTION
185.72
AI safety is still largely framed as alignment: training models to follow human preferences, safety policies, and normative constraints. That framing has improved the behavior of modern language models, but aligned behavior does not by itself guarantee that a deployed agent can be stopped, overridden, or constrained on...
[ -0.0022690000478178263, -0.019324999302625656, -0.09358099848031998, -0.033254001289606094, 0.05132000148296356, 0.09982399642467499, 0.04536600038409233, 0.05157700181007385, 0.00470300018787384, 0.06263700127601624, 0.02613699994981289, -0.0097850002348423, 0.08657299727201462, -0.039393...
[ -0.024265000596642494, -0.05521199852228165, -0.10194499790668488, 0.013477000407874584, 0.07544200122356415, 0.055470000952482224, -0.00023999999393709004, 0.006891000084578991, -0.0563259981572628, 0.031658999621868134, -0.019876999780535698, -0.039083000272512436, 0.04993100091814995, -...
AI Security, Red Teaming & Defense 2026
No code repository attached.
AI safety is still largely framed as alignment: training models to follow human preferences, safety policies, and normative constraints.
A system may be safe in expectation and still fail to yield to explicit runtime authority under conflicting instructions, long-horizon execution, adversarial inputs, or risky tool use.
Experiments with OpenClaw-based agents show that current alignment and guardrail mechanisms reduce risk, but often fail to provide persistent, authoritative, and enforceable runtime control.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:32:57.615658
2605.13471v1
Sleeper Channels and Provenance Gates: Persistent Prompt Injection in Always-on Autonomous AI Agents
AI Security, Red Teaming & Defense 2026
2026-05-13
[ "cs.CR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Always-on AI agents (OpenClaw, Hermes Agent) run as a single persistent process under the owner's identity, folding messaging, memory, self-authored skills, scheduling, and shell into one authority boundary. This configuration opens what we call \emp
Narek Maloyan
2
[ "Narek Maloyan", "Dmitry Namiot" ]
[]
http://arxiv.org/abs/2605.13471v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/maloyan/sleeper-channels", "https://github.com/nousresearch/hermes-agent", "https://github.com/openclaw/openclaw" ]
386,158
81,165
2026-08-13
NOASSERTION
185.07
Always-on AI agents (OpenClaw, Hermes Agent) run as a single persistent process under the owner's identity, folding messaging, memory, self-authored skills, scheduling, and shell into one authority boundary. This configuration opens what we call \emph{sleeper channels}: an untrusted input to one surface persists as a m...
[ -0.01912200078368187, -0.02450699917972088, -0.026319000869989395, 0.01863499917089939, 0.060210999101400375, -0.022630000486969948, 0.04877600073814392, -0.05324700102210045, 0.04299100115895271, 0.006430999841541052, -0.03427699953317642, -0.024661000818014145, 0.05878499895334244, 0.004...
[ -0.08927199989557266, -0.04039100185036659, -0.03743699938058853, -0.0051739998161792755, 0.05998900160193443, -0.04788700118660927, 0.03198999911546707, -0.0659480020403862, 0.060892000794410706, -0.014096000231802464, 0.005849000066518784, 0.004623000044375658, 0.035092998296022415, -0.0...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Always-on AI agents (OpenClaw, Hermes Agent) run as a single persistent process under the owner's identity, folding messaging, memory, self-authored skills, scheduling, and shell into one authority boundary.
This configuration opens what we call \emph{sleeper channels}: an untrusted input to one surface persists as a memory, skill, scheduled job, or filesystem patch, then fires later through a different surface with no attacker present.
Empirical evaluation is preregistered as follow-on.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:37:45.229548
2605.10779v1
LITMUS: Benchmarking Behavioral Jailbreaks of LLM Agents in Real OS Environments
AI Security, Red Teaming & Defense 2026
2026-05-11
[ "cs.CR", "cs.CL" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
The rapid proliferation of LLM-based autonomous agents in real operating system environments introduces a new category of safety risk beyond content safety: behavior jailbreak, where an adversary induces an agent to execute dangerous OS-level operati
Chiyu Zhang
11
[ "Chiyu Zhang", "Huiqin Yang", "Bendong Jiang", "Xiaolei Zhang", "Yiran Zhao", "Ruyi Chen", "Lu Zhou", "Xiaogang Xu", "Jiafei Wu", "Liming Fang", "Zhe Liu" ]
[]
http://arxiv.org/abs/2605.10779v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/AlienZhang1996/LITMUS", "https://github.com/openclaw/openclaw" ]
386,158
81,165
2026-08-13
NOASSERTION
184.97
The rapid proliferation of LLM-based autonomous agents in real operating system environments introduces a new category of safety risk beyond content safety: behavior jailbreak, where an adversary induces an agent to execute dangerous OS-level operations with irreversible consequences. Existing benchmarks either evaluat...
[ 0.018473999574780464, -0.03946699947118759, -0.02573000080883503, -0.05923999845981598, 0.042426999658346176, -0.05337899923324585, -0.03919700160622597, -0.06467299908399582, -0.052657999098300934, 0.01622300036251545, 0.030074000358581543, -0.048781998455524445, -0.03034999966621399, -0....
[ 0.015259999781847, -0.03278600051999092, -0.03894200176000595, -0.0176829993724823, 0.10044900327920914, -0.06067200005054474, -0.017566999420523643, -0.008449999615550041, -0.043912000954151154, 0.05008900165557861, 0.020167000591754913, -0.04947499930858612, 0.024470999836921692, 0.00344...
AI Security, Red Teaming & Defense 2026
No code repository attached.
The rapid proliferation of LLM-based autonomous agents in real operating system environments introduces a new category of safety risk beyond content safety: behavior jailbreak, where an adversary induces an agent to execute dangerous OS-level operations with irreversible consequences.
We present LITMUS (LLM-agents In-OS Testing for Measuring Unsafe Subversion), a benchmark addressing both gaps via a semantic-physical dual verification mechanism and OS-level state rollback.
Evaluation across frontier agents reveals three findings: (1) current agents lack effective safety awareness, with strong models (e.g., Claude Sonnet 4.6) still executing 40.64% of high-risk operations; (2) agents exhibit pervasive Execution Hallucination (EH), verbally refusing a request while the dangerous operation ...
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:38:52.867209
2604.23711v1
Spore: Efficient and Training-Free Privacy Extraction Attack on LLMs via Inference-Time Hybrid Probing
AI Security, Red Teaming & Defense 2026
2026-04-26
[ "cs.CR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
With the wide adoption of personal AI assistants such as OpenClaw, privacy leakage in user interaction contexts with large language model (LLM) agents has become a critical issue. Existing privacy attacks against LLMs primarily target training data,
Yu Cui
9
[ "Yu Cui", "Ruiqing Yue", "Hang Fu", "Sicheng Pan", "Zhuoyu Sun", "Baohan Huang", "Haibin Zhang", "Cong Zuo", "Licheng Wang" ]
[]
http://arxiv.org/abs/2604.23711v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/nousresearch/hermes-agent", "https://github.com/openclaw/openclaw" ]
386,158
81,165
2026-08-13
NOASSERTION
184.22
With the wide adoption of personal AI assistants such as OpenClaw, privacy leakage in user interaction contexts with large language model (LLM) agents has become a critical issue. Existing privacy attacks against LLMs primarily target training data, while research on inference-time contextual privacy risks in LLM agent...
[ -0.07616399973630905, -0.024971000850200653, 0.033511001616716385, 0.013811999931931496, 0.12026900053024292, -0.016899000853300095, 0.029203999787569046, -0.13791699707508087, 0.0612419992685318, 0.014198999851942062, 0.02269200049340725, 0.0051310001872479916, 0.0565590001642704, -0.0457...
[ -0.08033999800682068, -0.037629999220371246, -0.02459999918937683, 0.07502099871635437, 0.071042001247406, 0.031658001244068146, 0.07001399993896484, -0.049699001014232635, 0.05559699982404709, -0.02536500059068203, 0.022617999464273453, -0.06852000206708908, 0.07579399645328522, -0.032896...
AI Security, Red Teaming & Defense 2026
No code repository attached.
With the wide adoption of personal AI assistants such as OpenClaw, privacy leakage in user interaction contexts with large language model (LLM) agents has become a critical issue.
Moreover, prior methods often incur high attack costs, requiring multiple queries or relying on white-box assumptions, which limits their practicality in real-world deployments.
Our results show that \textsc{Spore} consistently bypasses both detection and strong safety alignment, demonstrating resilient performance in diverse defensive settings and real-world safety threats.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:45:00.371001
2606.16903v1
Directory-Aware Query and Maintenance in Vector Databases
Enterprise RAG & Vector Search 2026
2026-06-15
[ "cs.DB" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Vector databases typically manage metadata as flat scalar attributes, which limits their ability to express hierarchical directory semantics commonly used to organize code repositories, enterprise documents, and agent memories. As a result, directory
Mengzhao Wang
7
[ "Mengzhao Wang", "Zheng Gong", "Jingpei Hu", "Jiajie Fu", "Maojia Sheng", "Junwen Chen", "Yifan Zhu" ]
[]
http://arxiv.org/abs/2606.16903v1
VERIFIED_LIVE
https://github.com/nousresearch/hermes-agent
[ "https://github.com/nousresearch/hermes-agent", "https://github.com/infiniflow/ragflow", "https://github.com/anthropics/claude-code", "https://github.com/KurtPatrickHere/dir-vector-dataset" ]
229,907
0
Unknown
Unspecified
181.09
Vector databases typically manage metadata as flat scalar attributes, which limits their ability to express hierarchical directory semantics commonly used to organize code repositories, enterprise documents, and agent memories. As a result, directory-scoped retrieval and structural updates are often implemented as appl...
[ -0.006132999900728464, -0.0038479999639093876, -0.0036810000892728567, 0.04856700077652931, 0.04247700050473213, -0.06839100271463394, 0.06530600041151047, 0.013061000034213066, -0.05138099938631058, 0.001724999980069697, 0.0204709991812706, 0.06160400062799454, 0.012019000016152859, -0.02...
[ -0.016138000413775444, -0.005367000121623278, 0.012133999727666378, 0.024265000596642494, 0.006351000163704157, -0.099031001329422, -0.011648000217974186, 0.003957000095397234, -0.032976001501083374, -0.0006580000044777989, -0.000582000007852912, 0.03860300034284592, 0.05131300166249275, 0...
Enterprise RAG & Vector Search 2026
git clone https://github.com/nousresearch/hermes-agent
Vector databases typically manage metadata as flat scalar attributes, which limits their ability to express hierarchical directory semantics commonly used to organize code repositories, enterprise documents, and agent memories.
As a result, directory-scoped retrieval and structural updates are often implemented as application-layer workarounds, making recursive scope resolution expensive and directory maintenance difficult to keep consistent.
We then evaluate three implementation strategies: query-time path expansion (PE-Online), ingestion-time path expansion (PE-Offline), and a Trie-based Hierarchical Index (TrieHI).
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:20.849493
2605.08460v1
When Child Inherits: Modeling and Exploiting Subagent Spawn in Multi-Agent Networks
AI Security, Red Teaming & Defense 2026
2026-05-08
[ "cs.CR", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Since the official release of ChatGPT in 2022, large language models (LLMs) have rapidly evolved from chatbot-style interfaces into agentic systems that can delegate work through tools and newly spawned subagents. While these capabilities improve aut
Ziwen Cai
3
[ "Ziwen Cai", "Yihe Zhang", "Xiali Hei" ]
[]
http://arxiv.org/abs/2605.08460v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/agent0ai/agent-zero", "https://github.com/jgamblin/OpenClawCVEs", "https://github.com/nousresearch/hermes-agent" ]
229,896
8
2026-08-13
MIT
179.19
Since the official release of ChatGPT in 2022, large language models (LLMs) have rapidly evolved from chatbot-style interfaces into agentic systems that can delegate work through tools and newly spawned subagents. While these capabilities improve automation and scalability, they also pose new security risks in multi-ag...
[ -0.028053000569343567, -0.0147280003875494, -0.02994699962437153, -0.031184999272227287, 0.03456699848175049, -0.009952999651432037, 0.008923999965190887, 0.010083000175654888, 0.013206000439822674, 0.03251200169324875, 0.012198000214993954, -0.049872998148202896, 0.024650000035762787, 0.0...
[ -0.014183999970555305, -0.04420600086450577, -0.025851000100374222, -0.04326599836349487, 0.016339000314474106, 0.017362000420689583, 0.04730900004506111, -0.010892000049352646, 0.03016900084912777, -0.013469000346958637, 0.005809000227600336, -0.01040900032967329, 0.08036799728870392, 0.0...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Since the official release of ChatGPT in 2022, large language models (LLMs) have rapidly evolved from chatbot-style interfaces into agentic systems that can delegate work through tools and newly spawned subagents.
While these capabilities improve automation and scalability, they also pose new security risks in multi-agent networks.
Our findings show that inheritance is not merely an implementation detail, but a central component influencing the security of multi-agent systems.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:40:03.779029
2606.18356v1
SafeClawBench: Separating Semantic, Audit-Evidence, and Sandbox Harm in Tool-Using LLM Agents
AI Security, Red Teaming & Defense 2026
2026-06-16
[ "cs.CR", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Tool-using language-model agents introduce security failures that go beyond unsafe text: they can disclose protected objects, write persistent memory, send messages, modify databases, or trigger harmful code and tool effects. Existing evaluations oft
Yuchuan Tian
8
[ "Yuchuan Tian", "Mengyu Zheng", "Haocheng Mei", "Ye Yuan", "Chao Xu", "Xinghao Chen", "Hanting Chen", "Yu Wang" ]
[]
http://arxiv.org/abs/2606.18356v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/Significant-Gravitas/AutoGPT", "https://github.com/langchain-ai/langchain" ]
186,582
24,005
2026-08-13
MIT
178.87
Tool-using language-model agents introduce security failures that go beyond unsafe text: they can disclose protected objects, write persistent memory, send messages, modify databases, or trigger harmful code and tool effects. Existing evaluations often collapse these stages into a single attack success rate, making it ...
[ -0.01608400046825409, 0.026892000809311867, -0.050866998732089996, -0.022012999281287193, 0.11614199727773666, -0.00955200009047985, -0.022732999175786972, -0.03607799857854843, 0.03597399964928627, 0.036472998559474945, 0.027865000069141388, -0.03910600021481514, 0.07856100052595139, -0.0...
[ 0.013704000040888786, -0.0337660014629364, -0.05375700071454048, 0.03315899893641472, 0.08342500030994415, 0.018396999686956406, -0.033319998532533646, -0.020243000239133835, 0.009122000075876713, 0.003260999917984009, -0.017494000494480133, -0.0762609988451004, 0.10410899668931961, 0.0023...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Tool-using language-model agents introduce security failures that go beyond unsafe text: they can disclose protected objects, write persistent memory, send messages, modify databases, or trigger harmful code and tool effects.
We introduce SafeClawBench, a staged benchmark for tool-using agent security with 600 controlled adversarial tasks across six attack families: direct and indirect prompt injection, tool-return injection, memory poisoning, memory extraction, and ambiguity-driven unsafe inference.
Evaluating five agent endpoints under four prompt-level policies, we find that these endpoints capture different failure modes.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:25:27.052702
2606.15899v1
SkillVetBench: LLM-as-Judge for Multi-Dimensional Security Risk Evaluation in Open-Source LLM Agent Skills
AI Security, Red Teaming & Defense 2026
2026-06-14
[ "cs.CR", "cs.AI", "cs.HC", "cs.LG", "cs.MA" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Open-source LLM agent ecosystems are growing rapidly, yet the security of community-contributed skills - modular tool definitions that extend agent capabilities - remains largely unvetted. The gap we fill: existing scanners operate at the code layer
Ismail Hossain
5
[ "Ismail Hossain", "Sai Puppala", "Md Jahangir Alam", "Tanzim Ahad", "Sajedul Talukder" ]
[]
http://arxiv.org/abs/2606.15899v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/Significant-Gravitas/AutoGPT", "https://github.com/langchain-ai/langchain", "https://github.com/supreme-lab/SkillVetBench" ]
186,582
24,005
2026-08-13
MIT
178.77
Open-source LLM agent ecosystems are growing rapidly, yet the security of community-contributed skills - modular tool definitions that extend agent capabilities - remains largely unvetted. The gap we fill: existing scanners operate at the code layer and are structurally blind to instruction-layer and multi-agent risk -...
[ 0.020896999165415764, -0.009638999588787556, -0.10179299861192703, -0.025658000260591507, 0.055153001099824905, 0.032409001141786575, 0.04936699941754341, -0.0055669997818768024, -0.02014699950814247, -0.022873999550938606, -0.03721199929714203, -0.05082400143146515, 0.09614499658346176, 0...
[ 0.01688399910926819, -0.026976000517606735, -0.0825520008802414, -0.029357999563217163, 0.0717260017991066, -0.014929000288248062, 0.01243400014936924, -0.006486999802291393, -0.05359499901533127, -0.02553199976682663, -0.02628600038588047, -0.01840299926698208, 0.10599300265312195, 0.0289...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Open-source LLM agent ecosystems are growing rapidly, yet the security of community-contributed skills - modular tool definitions that extend agent capabilities - remains largely unvetted.
The gap we fill: existing scanners operate at the code layer and are structurally blind to instruction-layer and multi-agent risk - natural-language directives that hijack an agent, exfiltrate data through encoded side channels, or chain harm across pipelines - so what is needed is a semantic, multi-dimensional vetting...
What is demonstrated: drawing on our companion benchmark paper [ 1], the LLM-as-Judge stage achieves zero false negatives across 78 confirmed-malicious skills and zero false positives across 22 benign controls, while the best static baseline (SKILLSIEVE) still misses 15%; for instruction-layer categories such as Prompt...
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:25:56.947984
2606.14154v1
SkillMutator: Benchmarking and Defending Language-and-Code Cross-modal Attacks on LLM Agent Skills
AI Security, Red Teaming & Defense 2026
2026-06-12
[ "cs.CR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Large language model (LLM) agents increasingly extend their capabilities at runtime by loading Agent Skills, which pair natural-language specifications (SKILL.md) with executable scripts and resources. Because a skill's behavior relies on both natura
Youngduk Kim
3
[ "Youngduk Kim", "Minkyoo Song", "Seungwon Shin" ]
[]
http://arxiv.org/abs/2606.14154v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/protectai/llm-guard", "https://github.com/anthropics/skills", "https://github.com/snyk/agent-scan" ]
168,738
440
2026-07-08
MIT
177.58
Large language model (LLM) agents increasingly extend their capabilities at runtime by loading Agent Skills, which pair natural-language specifications (SKILL.md) with executable scripts and resources. Because a skill's behavior relies on both natural-language instructions and executable code, assessing its safety requ...
[ -0.08292900025844574, -0.04670799896121025, 0.0068660001270473, -0.02508299984037876, 0.018561000004410744, 0.022047000005841255, 0.05797199904918671, -0.0878399983048439, -0.02128699980676174, -0.06733299791812897, 0.03646999970078468, -0.08363299816846848, 0.05568600073456764, -0.0253870...
[ -0.04954399913549423, -0.07263399660587311, -0.01945200003683567, -0.005915000103414059, 0.08720699697732925, -0.03403700143098831, 0.036720000207424164, -0.029495999217033386, -0.01832900010049343, -0.04400099813938141, -0.0011889999732375145, -0.02513599954545498, 0.08126799762248993, 0....
AI Security, Red Teaming & Defense 2026
No code repository attached.
Large language model (LLM) agents increasingly extend their capabilities at runtime by loading Agent Skills, which pair natural-language specifications (SKILL.md) with executable scripts and resources.
Attackers can present a benign workflow in SKILL.md while embedding implicit directives that steer the agent to exfiltrate sensitive files, even if the scripts appear harmless.
These results show practical defense against cross-modal attacks is feasible without relying on costly frontier models.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:26:29.657603
2606.09563v1
PRISM: Recovering Instruction Sets from Language Model Activations
AI Security, Red Teaming & Defense 2026
2026-06-08
[ "cs.AI", "cs.LG" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
As LLMs are deployed as agents, reliable monitoring requires knowing not only what they output, but which instructions are steering their behavior. This is difficult when models infer unintended subgoals, follow contextual cues, or are influenced by
Gilad Gressel
6
[ "Gilad Gressel", "Rahul Pankajakshan", "Julia Diament", "Efim Hudis", "Krishnashree Achuthan", "Yisroel Mirsky" ]
[]
http://arxiv.org/abs/2606.09563v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/tatsu-lab/stanford_alpaca", "https://github.com/f/prompts.chat" ]
167,080
3,991
2024-07-17
Apache-2.0
177.27
As LLMs are deployed as agents, reliable monitoring requires knowing not only what they output, but which instructions are steering their behavior. This is difficult when models infer unintended subgoals, follow contextual cues, or are influenced by prompt injections and hidden objectives. While activation-to-language ...
[ -0.02399899996817112, -0.08328600227832794, 0.046199001371860504, 0.007087999954819679, 0.019576000049710274, 0.03136700019240379, 0.02463500015437603, 0.00017600000137463212, -0.024400999769568443, -0.05145899951457977, 0.010815000161528587, 0.005483999848365784, 0.020054999738931656, 0.0...
[ 0.020747000351548195, -0.024435000494122505, -0.0012959999730810523, 0.03994600102305412, 0.0609310008585453, 0.050321999937295914, 0.060224998742341995, -0.05958700180053711, 0.014310999773442745, -0.02855999954044819, -0.020963000133633614, -0.058097999542951584, 0.08868099749088287, -0....
AI Security, Red Teaming & Defense 2026
No code repository attached.
As LLMs are deployed as agents, reliable monitoring requires knowing not only what they output, but which instructions are steering their behavior.
While activation-to-language methods suggest that hidden states can reveal natural-language information, existing approaches are not designed to recover the full set of simultaneous instructions, constraints, prohibitions, and subgoals active in agentic settings.
Across benign, constrained, prompt-injection, and hidden-objective settings, PRISM outperforms activation-to-language baselines, especially on security-relevant objectives.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:27:45.543030
2607.03423v1
Securing Multi-Tool AI Agent Chains With Dynamic, Real-Time Compositional Policies
AI Security, Red Teaming & Defense 2026
2026-07-03
[ "cs.CR", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Modern AI agent implementations such as frontier coding agents chain multiple tools at runtime that create a security surface that per-tool guardrails are unable to address, as individually permitted tools can violate organizational policies when com
Chris Schneider
4
[ "Chris Schneider", "Kriti Faujdar", "Philipp Schoenegger", "Ben Bariach" ]
[]
http://arxiv.org/abs/2607.03423v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/langchain-ai/langchain", "https://github.com/features/copilot" ]
144,158
24,005
2026-08-13
MIT
176.92
Modern AI agent implementations such as frontier coding agents chain multiple tools at runtime that create a security surface that per-tool guardrails are unable to address, as individually permitted tools can violate organizational policies when composed. We propose the Dynamic Security Control Compositor (DSCC), a tw...
[ -0.08087100088596344, 0.022099999710917473, -0.051396001130342484, -0.005967000033706427, -0.03447800129652023, -0.0050050001591444016, -0.008040999993681908, -0.07135199755430222, 0.04360299929976463, 0.033851999789476395, -0.04257600009441376, -0.05526699870824814, 0.06158199906349182, -...
[ -0.029496999457478523, 0.007393000181764364, -0.046929001808166504, -0.010528000071644783, 0.029510000720620155, -0.01779700070619583, 0.031261999160051346, -0.06523499637842178, 0.0021059999708086252, 0.015355999581515789, -0.05556200072169304, -0.07064799964427948, 0.04960300028324127, -...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Modern AI agent implementations such as frontier coding agents chain multiple tools at runtime that create a security surface that per-tool guardrails are unable to address, as individually permitted tools can violate organizational policies when composed.
We propose the Dynamic Security Control Compositor (DSCC), a two-phase approach to compositional security for multi-tool agent chains.
We provide a reference implementation on 32 tools governed by 16 NIST SP 800-53 aligned policies and evaluate it under two composition modes.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:22:13.922410
2606.19458v1
MonaVec: A Training-Free Embedded Vector Search Kernel for Edge and Offline AI Systems
Enterprise RAG & Vector Search 2026
2026-06-17
[ "cs.IR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
We present MonaVec, a deterministic, embedded vector-search kernel for edge and offline AI -- settings where server infrastructure, network connectivity, and training data are all unavailable. Existing vector-search systems assume a persistent server
Oğuzhan Yenen
1
[ "Oğuzhan Yenen" ]
[]
http://arxiv.org/abs/2606.19458v1
VERIFIED_LIVE
https://github.com/ggerganov/llama.cpp
[ "https://github.com/unum-cloud/usearch", "https://github.com/mlc-ai/mlc-llm", "https://github.com/ggerganov/llama.cpp", "https://github.com/mona-hq/monavec-edge-bench" ]
123,767
0
Unknown
Unspecified
174.47
We present MonaVec, a deterministic, embedded vector-search kernel for edge and offline AI -- settings where server infrastructure, network connectivity, and training data are all unavailable. Existing vector-search systems assume a persistent server, gigabytes of RAM, or a training pass over the corpus; MonaVec instea...
[ -0.05371199920773506, -0.08365599811077118, -0.034265000373125076, 0.04185299947857857, 0.10650400072336197, -0.013125999830663204, -0.05114699900150299, 0.006165999919176102, -0.0942389965057373, -0.05665599927306175, -0.014090999960899353, -0.04931100085377693, -0.020641999319195747, -0....
[ -0.04050299897789955, -0.06509599834680557, -0.04927600175142288, 0.013505999930202961, -0.005534000229090452, -0.04205600172281265, -0.059842001646757126, 0.012616000138223171, -0.057677000761032104, -0.018797000870108604, -0.007703999988734722, -0.021316999569535255, 0.0015160000184550881,...
Enterprise RAG & Vector Search 2026
git clone https://github.com/ggerganov/llama.cpp
We present MonaVec, a deterministic, embedded vector-search kernel for edge and offline AI -- settings where server infrastructure, network connectivity, and training data are all unavailable.
Existing vector-search systems assume a persistent server, gigabytes of RAM, or a training pass over the corpus; MonaVec instead targets the deployment profile of SQLite: one file, one function call, runs anywhere.
The index persists as a single .mvec file whose embedded ChaCha20 rotation seed makes results reproducible across architectures and byte-identical within a build -- a determinism guarantee that parallel-build graph libraries cannot offer.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:20.715744
2606.13126v1
MiniPIC: Flexible Position-Independent Caching in <100LOC
Enterprise RAG & Vector Search 2026
2026-06-11
[ "cs.LG", "cs.AI", "cs.CL" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Retrieval-augmented and agentic workloads repeatedly prefill recurring predictable structured inputs (which we call "spans") such as documents and code files. Yet, prefix caching in engines such as vLLM cannot reuse their KV entries unless they share
Nathan Ordonez
2
[ "Nathan Ordonez", "Thomas Parnell" ]
[ "IBM Research" ]
http://arxiv.org/abs/2606.13126v1
VERIFIED_LIVE
https://github.com/ggerganov/llama.cpp
[ "https://github.com/NVIDIA/TensorRT-LLM", "https://github.com/meta-llama/llama3", "https://github.com/ggerganov/llama.cpp", "https://github.com/huggingface/transformers" ]
123,767
0
Unknown
Unspecified
174.17
Retrieval-augmented and agentic workloads repeatedly prefill recurring predictable structured inputs (which we call "spans") such as documents and code files. Yet, prefix caching in engines such as vLLM cannot reuse their KV entries unless they share identical prefixes with another request, while Position-Independent C...
[ -0.06000899896025658, 0.03663399815559387, -0.004153000190854073, 0.009859000332653522, 0.017256999388337135, -0.0768750011920929, 0.028860999271273613, -0.000577000027988106, 0.016536999493837357, 0.05548899993300438, -0.020732000470161438, 0.053720999509096146, -0.009695000015199184, -0....
[ -0.06230499967932701, 0.04434100165963173, 0.0092569999396801, 0.0227269995957613, -0.040049999952316284, -0.05270100012421608, -0.08035299926996231, -0.015440000221133232, 0.05459799990057945, -0.009492999874055386, -0.003914000000804663, 0.029812999069690704, -0.010382999666035175, -0.05...
Enterprise RAG & Vector Search 2026
git clone https://github.com/ggerganov/llama.cpp
Retrieval-augmented and agentic workloads repeatedly prefill recurring predictable structured inputs (which we call "spans") such as documents and code files.
We present Minimalistic PIC (MiniPIC): a minimal, flexible and fast vLLM design built from two ingredients: positional-encoding-free KV cache and user-controlled cache-reuse primitives.
On 2WikiMultihopQA, MiniPIC with interleaved scheduling improves prefill throughput by 49% over baseline vLLM, reduces cached-span time-to-first-token by up to two orders of magnitude, preserves the linear prefill scaling of uncached spans, and incurs only 5.7% worst-case overhead.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:24.159937
2606.11257v1
Energy-Efficient On-Device RAG on a Mobile NPU: System Design and Benchmark on Snapdragon X Elite
Enterprise RAG & Vector Search 2026
2026-06-09
[ "cs.CL", "cs.LG", "cs.PF" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Retrieval-Augmented Generation (RAG) pipelines are compute-intensive, combining embedding, retrieval, reranking, and large language model (LLM) generation. Running them entirely on-device benefits privacy, latency, and offline use, but the energy cos
Zhiyuan Cheng
2
[ "Zhiyuan Cheng", "Longying Lai" ]
[]
http://arxiv.org/abs/2606.11257v1
VERIFIED_LIVE
https://github.com/ggerganov/llama.cpp
[ "https://github.com/pytorch/executorch", "https://github.com/ggerganov/llama.cpp", "https://github.com/run-llama/llama_index", "https://github.com/langchain-ai/langchain" ]
123,767
0
Unknown
Unspecified
174.07
Retrieval-Augmented Generation (RAG) pipelines are compute-intensive, combining embedding, retrieval, reranking, and large language model (LLM) generation. Running them entirely on-device benefits privacy, latency, and offline use, but the energy cost of CPU inference is a major barrier. We present what is, to our know...
[ -0.06159599870443344, 0.08376699686050415, 0.05456199869513512, -0.010053999722003937, -0.0013119999784976244, -0.1578650027513504, -0.02923399955034256, -0.011839999817311764, -0.08769799768924713, -0.008837999776005745, -0.0204709991812706, -0.01733200065791607, 0.03488999977707863, -0.0...
[ -0.06496699899435043, -0.020306000486016273, -0.017240000888705254, 0.04212699830532074, 0.019317999482154846, -0.048833999782800674, -0.0574599988758564, 0.00020300000323913991, -0.014821999706327915, -0.07325299829244614, -0.04813100025057793, -0.006701999809592962, -0.0064500002190470695,...
Enterprise RAG & Vector Search 2026
git clone https://github.com/ggerganov/llama.cpp
Retrieval-Augmented Generation (RAG) pipelines are compute-intensive, combining embedding, retrieval, reranking, and large language model (LLM) generation.
We present what is, to our knowledge, the first end-to-end RAG pipeline that runs all neural stages -- embedding, reranking, and LLM generation -- on the Qualcomm Hexagon NPU of the Snapdragon X Elite.
On indexing, the NPU achieves 9.1x higher embedding throughput and 12.3x less system energy.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:22.183982
2606.05958v1
Steering Vectors are an Adversarial Attack Surface
AI Security, Red Teaming & Defense 2026
2026-06-04
[ "cs.LG" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Activation steering has become a popular way to control Large Language Model (LLM) behavior without fine-tuning. Since the technique is plug-and-play, users share datasets and precomputed vectors to steer model activations. However, we show that a \e
Abzal Aidakhmetov
6
[ "Abzal Aidakhmetov", "Donato Crisostomi", "Tommaso Mencattini", "Adrian Robert Minut", "Iacopo Masi", "Emanuele Rodolà" ]
[]
http://arxiv.org/abs/2606.05958v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/AbzalAidakhmetov/adversarial_attack", "https://github.com/steering-vectors/steering-vectors", "https://github.com/ggml-org/llama.cpp" ]
123,763
0
Unknown
Unspecified
173.81
Activation steering has become a popular way to control Large Language Model (LLM) behavior without fine-tuning. Since the technique is plug-and-play, users share datasets and precomputed vectors to steer model activations. However, we show that a \emph{stealth data poisoning attack} silently compromises this pipeline....
[ -0.04330199956893921, -0.0778689980506897, -0.009394999593496323, 0.044512998312711716, 0.11109499633312225, 0.009513000026345253, 0.049587998539209366, -0.05548999831080437, 0.0048500001430511475, -0.010162999853491783, -0.002025000052526593, -0.019300000742077827, 0.06481000036001205, -0...
[ -0.017196999862790108, -0.07108300179243088, 0.05074800178408623, -0.0009919999865815043, 0.05721300095319748, -0.016016000881791115, 0.06381399929523468, -0.012811999768018723, 0.05707700178027153, 0.018370000645518303, 0.07067299634218216, -0.003913000226020813, 0.10132600367069244, -0.0...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Activation steering has become a popular way to control Large Language Model (LLM) behavior without fine-tuning.
Since the technique is plug-and-play, users share datasets and precomputed vectors to steer model activations.
Finally, we find that a refusal-direction orthogonalization defense can recover ${\approx}82\%$ of the ASR gap without harming benign behavior.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:28:51.680277
2605.19722v1
Measuring Safety Alignment Effects in Autonomous Security Agents
AI Security, Red Teaming & Defense 2026
2026-05-19
[ "cs.CR", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Do stock safety-aligned language models and their uncensored or abliterated derivatives behave differently when run as autonomous security agents? Single-turn refusal benchmarks cannot answer this question: security agents must inspect repositories,
Isaac David
2
[ "Isaac David", "Arthur Gervais" ]
[]
http://arxiv.org/abs/2605.19722v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/ggerganov/llama.cpp" ]
123,764
0
Unknown
Unspecified
173.01
Do stock safety-aligned language models and their uncensored or abliterated derivatives behave differently when run as autonomous security agents? Single-turn refusal benchmarks cannot answer this question: security agents must inspect repositories, call tools, and produce vulnerability evidence inside authorized sandb...
[ 0.0070500001311302185, -0.05434500053524971, -0.13162000477313995, -0.08173099905252457, 0.06374099850654602, 0.00874600000679493, 0.05660799890756607, -0.012600000016391277, 0.034970998764038086, 0.03549300134181976, 0.04431600123643875, -0.033608999103307724, 0.07507000118494034, -0.0010...
[ -0.018471000716090202, -0.028907999396324158, -0.006994999945163727, -0.041868001222610474, 0.017014000564813614, -0.04961400106549263, -0.03310300037264824, -0.017565999180078506, 0.023156000301241875, -0.0020439999643713236, 0.028085999190807343, -0.05270799994468689, 0.045600999146699905,...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Do stock safety-aligned language models and their uncensored or abliterated derivatives behave differently when run as autonomous security agents?
We present a trace-based benchmark of 30 local vulnerability-analysis tasks with fixed tools, deterministic success predicates, redaction rules, and grounding checks, and compare four stock models against uncensored or abliterated derivatives: Gemma 4 31B, Gemma 4 26B A4B, Qwen2.5-Coder 7B, and Llama 3.1 8B.
These results show that safety alignment effects in autonomous security agents should be measured at the system level, separating refusal, unsafe action, tool reliability, and evidence grounding rather than treating refusal rate as the safety signal.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:35:09.644145
2605.07068v1
WiCER: Wiki-memory Compile, Evaluate, Refine Iterative Knowledge Compilation for LLM Wiki Systems
Enterprise RAG & Vector Search 2026
2026-05-08
[ "cs.CL", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
The LLM Wiki pattern, to compile and provide domain knowledge into a persistent artifact and serve it to LLMs via KV cache inference, promises context access at sub-second latency with zero retrieval failure. Realizing this requires solving the compi
Juan M. Huerta
1
[ "Juan M. Huerta" ]
[]
http://arxiv.org/abs/2605.07068v1
VERIFIED_LIVE
https://github.com/ggml-org/llama.cpp
[ "https://github.com/ggml-org/llama.cpp" ]
123,767
0
Unknown
Unspecified
172.47
The LLM Wiki pattern, to compile and provide domain knowledge into a persistent artifact and serve it to LLMs via KV cache inference, promises context access at sub-second latency with zero retrieval failure. Realizing this requires solving the compilation gap: LLM compilation distilling raw documents into a wiki witho...
[ -0.0057830000296235085, 0.00571299996227026, -0.016047999262809753, -0.019334999844431877, 0.04924200102686882, -0.0903249979019165, 0.005481999833136797, 0.03910300135612488, -0.018232999369502068, 0.04308899864554405, 0.0013200000394135714, 0.016600999981164932, 0.06520500034093857, 0.00...
[ -0.04761099815368652, -0.010635999962687492, 0.020989999175071716, 0.047384001314640045, 0.0651020035147667, -0.09226500242948532, -0.002882000058889389, 0.03675999864935875, 0.006810000166296959, -0.0026859999634325504, 0.004286999814212322, -0.03479800000786781, 0.04946500062942505, -0.0...
Enterprise RAG & Vector Search 2026
git clone https://github.com/ggml-org/llama.cpp
The LLM Wiki pattern, to compile and provide domain knowledge into a persistent artifact and serve it to LLMs via KV cache inference, promises context access at sub-second latency with zero retrieval failure.
To address the compilation gap, we propose WiCER (Wiki-memory Compile, Evaluate, Refine), an iterative algorithm inspired by counterexample-guided abstraction refinement (CEGAR) that closes this gap.
WiCER evaluates compiled wikis against diagnostic probes, identifies dropped facts, and forces their preservation in subsequent compilations.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:54.124884
2606.07992v1
VATS: Exploiting Implicit Authority in Error-Path Injection via Systematic Mutation
AI Security, Red Teaming & Defense 2026
2026-06-06
[ "cs.AI", "cs.CR", "cs.SE" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
As the Model Context Protocol (MCP) standardizes tool-calling for autonomous agents, it introduces a critical, unexamined attack surface: the error-handling loop. We hypothesize that tool error messages possess implicit authority, triggering correcti
Harshil Patel
2
[ "Harshil Patel", "Kunal Pai" ]
[]
http://arxiv.org/abs/2606.07992v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/openai/codex", "https://github.com/google-gemini/gemini-cli" ]
106,497
0
Unknown
Unspecified
172.28
As the Model Context Protocol (MCP) standardizes tool-calling for autonomous agents, it introduces a critical, unexamined attack surface: the error-handling loop. We hypothesize that tool error messages possess implicit authority, triggering corrective reasoning modes that bypass standard safety heuristics. We introduc...
[ -0.05182399973273277, 0.07869499921798706, 0.02642500028014183, 0.009526999667286873, -0.054019998759031296, -0.0910789966583252, 0.00559299997985363, 0.015348000451922417, -0.022776000201702118, -0.0044450000859797, 0.05131400004029274, 0.0010969999711960554, -0.012505999766290188, -0.016...
[ -0.042608000338077545, -0.024108000099658966, -0.02545199915766716, 0.01259199995547533, 0.05741100013256073, -0.05340399965643883, 0.022161999717354774, -0.01540399994701147, 0.011535000056028366, 0.008907999843358994, -0.01855899952352047, -0.0736989974975586, 0.06538499891757965, 0.0132...
AI Security, Red Teaming & Defense 2026
No code repository attached.
As the Model Context Protocol (MCP) standardizes tool-calling for autonomous agents, it introduces a critical, unexamined attack surface: the error-handling loop.
We introduce VATS (Vulnerability Analysis of Tool Streams), a mutation-driven framework that systematically evolves adversarial payloads across seven structural and linguistic dimensions.
While we find that production framework guardrails can mitigate these vulnerabilities, the inherent susceptibility of the model layer poses a systemic risk to bespoke agentic workflows.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:28:19.831994
2607.03440v1
Improving Access to Historical Archives with Real-time RAG-based Systems
Enterprise RAG & Vector Search 2026
2026-07-03
[ "cs.IR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Digitized historical archives are large, heterogeneous cultural heritage repositories, but access methods for such archives face challenges such as noisy optical character recognition (OCR) output and rigid keyword-based retrieval, which limit retrie
Stergios Konstantinidis
6
[ "Stergios Konstantinidis", "Hayman Lotfy", "Alexis Erne", "Faruk Zahiragic", "Min-Yen Kan", "Michalis Vlachos" ]
[]
http://arxiv.org/abs/2607.03440v1
VERIFIED_LIVE
https://github.com/deepseek-ai/DeepSeek-R1
[ "https://github.com/Stergios-Konstantinidis/Improving-Access-to-Vast-Historical-Archives-with-Large-Language-Models", "https://github.com/deepseek-ai/DeepSeek-R1", "https://github.com/JaidedAI/EasyOCR" ]
91,973
11,707
2025-12-05
Apache-2.0
172.04
Digitized historical archives are large, heterogeneous cultural heritage repositories, but access methods for such archives face challenges such as noisy optical character recognition (OCR) output and rigid keyword-based retrieval, which limit retrieval quality. In this work, we present an end-to-end archival processin...
[ -0.08867199718952179, 0.10043299943208694, -0.07295800000429153, -0.04111599922180176, -0.0530799999833107, -0.00914900004863739, -0.11431299895048141, -0.010854999534785748, -0.022944999858736992, 0.025458000600337982, -0.03404400125145912, 0.096220001578331, 0.03692600131034851, -0.03297...
[ -0.06488899886608124, 0.0002640000020619482, 0.01577799953520298, 0.017880000174045563, 0.04639500007033348, 0.019262999296188354, -0.08061300218105316, 0.0013889999827370048, -0.002894999925047159, -0.023569999262690544, -0.009134000167250633, 0.04179999977350235, 0.047123998403549194, 0....
Enterprise RAG & Vector Search 2026
git clone https://github.com/deepseek-ai/DeepSeek-R1
Digitized historical archives are large, heterogeneous cultural heritage repositories, but access methods for such archives face challenges such as noisy optical character recognition (OCR) output and rigid keyword-based retrieval, which limit retrieval quality.
In this work, we present an end-to-end archival processing and retrieval framework that integrates large language models (LLMs) into the archival pipeline.
These results demonstrate that integrating LLMs with established document processing and retrieval pipelines can elevate digital libraries from static repositories to interactive, semantically searchable archival systems.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:02.290714
2605.28122v1
SNARE: Adaptive Scenario Synthesis for Eliciting Overeager Behavior in Coding Agents
AI Security, Red Teaming & Defense 2026
2026-05-27
[ "cs.CR", "cs.AI", "cs.CL" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
A coding agent executes a benign task as a sequence of shell, file, and network actions, any of which can quietly exceed the authorized scope while the task still completes. We call this overeager behavior: the prompt is not adversarial and the run s
Yubin Qu
7
[ "Yubin Qu", "Yi Liu", "Gelei Deng", "Yanjun Zhang", "Yuekang Li", "Ying Zhang", "Leo Yu Zhang" ]
[]
http://arxiv.org/abs/2605.28122v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/BerriAI/litellm", "https://github.com/openai/codex", "https://github.com/google-gemini/gemini-cli" ]
106,497
0
Unknown
Unspecified
171.78
A coding agent executes a benign task as a sequence of shell, file, and network actions, any of which can quietly exceed the authorized scope while the task still completes. We call this overeager behavior: the prompt is not adversarial and the run succeeds, yet an out-of-scope step can leak credentials or delete files...
[ 0.007836000062525272, -0.09522800147533417, 0.019703999161720276, 0.00019299999985378236, 0.009688000194728374, -0.014353999868035316, 0.029681000858545303, 0.004885000176727772, 0.005154999904334545, 0.011958999559283257, -0.027698000892996788, -0.08136799931526184, -0.013435999862849712, ...
[ -0.051180001348257065, -0.03216199949383736, -0.043083999305963516, 0.01870100013911724, 0.05510199815034866, -0.03639199957251549, -0.005001000128686428, -0.04898900166153908, 0.014453000389039516, 0.021976999938488007, -0.04427700117230415, -0.01989700086414814, 0.009630000218749046, -0....
AI Security, Red Teaming & Defense 2026
No code repository attached.
A coding agent executes a benign task as a sequence of shell, file, and network actions, any of which can quietly exceed the authorized scope while the task still completes.
We present SNARE (Synthesizing Non-adversarial scenarios for Adaptive Reward-guided Elicitation), a pipeline that composes benign scenarios from reusable scope and trap fragments, scores each run with a judge-free oracle flagging trap-pattern matches and unsolicited file additions or deletions, and uses Thompson sampli...
This variation is driven by the agent framework, not the model: the framework accounts for 56% of it against the model's 21%, so any single-framework or single-model evaluation undercounts the matrix by about a fifth.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:32:35.874063
2605.18583v1
Overeager Coding Agents: Measuring Out-of-Scope Actions on Benign Tasks
AI Security, Red Teaming & Defense 2026
2026-05-18
[ "cs.SE", "cs.AI", "cs.CL", "cs.CR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Coding agents now run autonomously with shell, file, and network privileges. When a user issues a benign request, the agent sometimes does more than asked: it deletes unrelated files, wipes a stale credentials backup, or rewrites configuration the us
Yubin Qu
7
[ "Yubin Qu", "Ying Zhang", "Yanjun Zhang", "Gelei Deng", "Yuekang Li", "Leo Yu Zhang", "Yi Liu" ]
[]
http://arxiv.org/abs/2605.18583v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/All-Hands-AI/OpenHands", "https://github.com/openai/codex" ]
105,676
0
Unknown
Unspecified
171.25
Coding agents now run autonomously with shell, file, and network privileges. When a user issues a benign request, the agent sometimes does more than asked: it deletes unrelated files, wipes a stale credentials backup, or rewrites configuration the user never mentioned. We call these scope expansions overeager actions, ...
[ 0.009285000152885914, -0.05721599981188774, -0.013918000273406506, 0.004786999896168709, 0.07409100234508514, -0.006118999794125557, 0.09426800161600113, 0.003909999970346689, -0.02179899998009205, 0.05576299875974655, 0.016878999769687653, -0.0001049999991664663, -0.034919001162052155, 0....
[ -0.005371000152081251, 0.012071999721229076, -0.031127000227570534, -0.03374699875712395, 0.013663000427186489, -0.06365100294351578, -0.010611999779939651, 0.014322999864816666, -0.011563999578356743, 0.0270990002900362, -0.025049999356269836, 0.00015300000086426735, 0.013852999545633793, ...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Coding agents now run autonomously with shell, file, and network privileges.
We present OverEager-Gen, a benchmark dedicated to overeager behavior on benign tasks.
Within-framework base-model variance reaches 15.9 pp, indicating that model-layer alignment does not fully propagate through permissive permission gating.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:35:38.608567
2605.12493v1
LongMemEval-V2: Evaluating Long-Term Agent Memory Toward Experienced Colleagues
Enterprise RAG & Vector Search 2026
2026-05-12
[ "cs.CL" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Long-term memory is crucial for agents in specialized web environments, where success depends on recalling interface affordances, state dynamics, workflows, and recurring failure modes. However, existing memory benchmarks for agents mostly focus on u
Di Wu
7
[ "Di Wu", "Zixiang Ji", "Asmi Kawatkar", "Bryan Kwan", "Jia-Chen Gu", "Nanyun Peng", "Kai-Wei Chang" ]
[]
http://arxiv.org/abs/2605.12493v1
VERIFIED_LIVE
https://github.com/openai/codex
[ "https://github.com/gkamradt/LLMTest_NeedleInAHaystack", "https://github.com/openai/codex", "https://github.com/ServiceNow/AgentLab" ]
105,680
0
Unknown
Unspecified
170.95
Long-term memory is crucial for agents in specialized web environments, where success depends on recalling interface affordances, state dynamics, workflows, and recurring failure modes. However, existing memory benchmarks for agents mostly focus on user histories, short traces, or downstream task success, leaving open ...
[ 0.0729999989271164, 0.0030249999836087227, -0.04392300173640251, 0.04443899914622307, 0.05178000032901764, 0.08346900343894958, -0.051660001277923584, 0.026325000450015068, -0.0015059999423101544, -0.06575000286102295, -0.016899999231100082, 0.030990000814199448, 0.0033950000070035458, 0.0...
[ 0.03264399990439415, -0.07455399632453918, -0.05454900115728378, 0.04597700014710426, 0.072223000228405, 0.020006999373435974, -0.018092000856995583, 0.012602999806404114, -0.05823799967765808, -0.0482419990003109, -0.04296699911355972, -0.042291998863220215, 0.06317099928855896, 0.0091580...
Enterprise RAG & Vector Search 2026
git clone https://github.com/openai/codex
Long-term memory is crucial for agents in specialized web environments, where success depends on recalling interface affordances, state dynamics, workflows, and recurring failure modes.
However, existing memory benchmarks for agents mostly focus on user histories, short traces, or downstream task success, leaving open how to directly evaluate whether memory systems effectively internalize environment-specific experience.
Together, these results establish LME-V2 as a challenging testbed for developing long-term memory systems for environment experience.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:51.444319
2605.16462v1
Asking Back: Interaction-Layer Antidistillation Watermarks
AI Security, Red Teaming & Defense 2026
2026-05-15
[ "cs.CR", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Detecting unauthorized knowledge distillation from a deployed LLM API is hard because the defender controls neither the attacker's training pipeline nor the next-token logits. Existing defenses operate on the teacher's output tokens -- biasing the ne
Guang Yang
6
[ "Guang Yang", "Amir Ghasemian", "Fengchen Liu", "Zhong Wang", "Ninareh Mehrabi", "Homa Hosseinmardi" ]
[]
http://arxiv.org/abs/2605.16462v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/vllm-project/vllm", "https://github.com/tatsu-lab/stanford_alpaca" ]
88,953
3,991
2024-07-17
Apache-2.0
169.23
Detecting unauthorized knowledge distillation from a deployed LLM API is hard because the defender controls neither the attacker's training pipeline nor the next-token logits. Existing defenses operate on the teacher's output tokens -- biasing the next-token distribution (green-list watermarks, cryptographic schemes, a...
[ -0.08868899941444397, 0.05024699866771698, 0.08699499815702438, 0.039788998663425446, 0.1341090053319931, -0.08513899892568588, 0.09183000028133392, -0.02222600020468235, -0.014814999885857105, -0.069022998213768, 0.0026400000788271427, -0.018943000584840775, 0.006839999929070473, 0.029965...
[ -0.05684499815106392, 0.01843699999153614, 0.03565400093793869, -0.002429000101983547, 0.09763599932193756, -0.01945200003683567, 0.05703900009393692, -0.013553000055253506, 0.01129399985074997, -0.02386000007390976, 0.017177000641822815, -0.019996000453829765, 0.07962799817323685, 0.00455...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Detecting unauthorized knowledge distillation from a deployed LLM API is hard because the defender controls neither the attacker's training pipeline nor the next-token logits.
Existing defenses operate on the teacher's output tokens -- biasing the next-token distribution (green-list watermarks, cryptographic schemes, antidistillation sampling) or rewriting outputs after generation.
The interaction layer is a viable design locus for antidistillation watermarking, complementary to token-, model-, and reasoning-trace-layer defenses.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:37:03.945331
2605.11217v1
Leveraging RAG for Training-Free Alignment of LLMs
Enterprise RAG & Vector Search 2026
2026-05-11
[ "cs.LG", "cs.AI", "cs.CR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Large language model (LLM) alignment algorithms typically consist of post-training over preference pairs. While such algorithms are widely used to enable safety guardrails and align LLMs with general human preferences, we show that state-of-the-art a
John T. Halloran
1
[ "John T. Halloran" ]
[]
http://arxiv.org/abs/2605.11217v1
VERIFIED_LIVE
https://github.com/modelcontextprotocol/servers
[ "https://github.com/modelcontextprotocol/servers", "https://github.com/stripe/agent-toolkit", "https://github.com/stevie1023/OPAD", "https://github.com/philschmid/mcp-openai-gemini-llama-example" ]
89,531
0
Unknown
Unspecified
169.1
Large language model (LLM) alignment algorithms typically consist of post-training over preference pairs. While such algorithms are widely used to enable safety guardrails and align LLMs with general human preferences, we show that state-of-the-art alignment algorithms require significant computational resources while ...
[ -0.0489719994366169, -0.03973599895834923, 0.026972999796271324, -0.10807999968528748, -0.08781000226736069, -0.014398000203073025, -0.046900998800992966, -0.011349000036716461, 0.015961000695824623, -0.0056420001201331615, 0.016378000378608704, 0.015201999805867672, 0.02284100092947483, -...
[ -0.03859499841928482, -0.00892299972474575, -0.03395000100135803, -0.060350000858306885, 0.03485099971294403, 0.06588900089263916, 0.02814899943768978, -0.02875800058245659, 0.10174699872732162, 0.00009899999713525176, 0.03689299896359444, -0.03956099972128868, 0.0561240017414093, 0.036187...
Enterprise RAG & Vector Search 2026
git clone https://github.com/modelcontextprotocol/servers
Large language model (LLM) alignment algorithms typically consist of post-training over preference pairs.
While such algorithms are widely used to enable safety guardrails and align LLMs with general human preferences, we show that state-of-the-art alignment algorithms require significant computational resources while being far less capable of enabling refusal guardrails for recent agentic attacks.
RAG-Pref is online (training-free), compatible with off-the-shelf packages, and, when combined with offline (training-based) alignment algorithms, enables more than an average 3.7 factor improvement in agentic attack refusals across five widely used LLMs, compared to 2.9 for other online alignment algorithms and 1.5 fo...
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:54.891590
2604.22871v1
AutoRISE: Agent-Driven Strategy Evolution for Red-Teaming Large Language Models
AI Security, Red Teaming & Defense 2026
2026-04-23
[ "cs.CR", "cs.AI", "cs.MA" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Automated red-teaming methods for large language models typically optimize attack prompts within a fixed, human-designed strategy, leaving the attack strategy itself unchanged. We instead optimize the strategy. We propose AutoRISE, a method that sear
Tanmay Gautam
3
[ "Tanmay Gautam", "Alireza Bahramali", "Sandeep Atluri" ]
[]
http://arxiv.org/abs/2604.22871v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/karpathy/autoresearch", "https://github.com/Tele-EVOL/TeleAI-Safety" ]
93,784
0
Unknown
Unspecified
168.7
Automated red-teaming methods for large language models typically optimize attack prompts within a fixed, human-designed strategy, leaving the attack strategy itself unchanged. We instead optimize the strategy. We propose AutoRISE, a method that searches over executable attack programs rather than individual prompts. A...
[ 0.033181000500917435, -0.0782570019364357, -0.04159900173544884, 0.030519999563694, 0.010634000413119793, 0.0906430035829544, 0.03860300034284592, 0.0005210000090301037, -0.014156999997794628, -0.06429299712181091, -0.047818999737501144, -0.05266200006008148, 0.04171599820256233, 0.0364530...
[ 0.009550999850034714, -0.02695300057530403, -0.04332200065255165, 0.03494499996304512, 0.00967200007289648, 0.06066500023007393, 0.04300500079989433, -0.0024649999104440212, -0.02578200027346611, -0.01691400073468685, -0.059014998376369476, -0.06753899902105331, 0.05037999898195267, 0.0099...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Automated red-teaming methods for large language models typically optimize attack prompts within a fixed, human-designed strategy, leaving the attack strategy itself unchanged.
We propose AutoRISE, a method that searches over executable attack programs rather than individual prompts.
Across held-out models, AutoRISE improves average attack success rate by 17.0 points over the strongest baseline, and improves attack success by up to 16 points on frontier targets with low baseline success rates.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:45:43.296824
2606.28370v1
Conversational Query Engine for Mixed-Modality Heterogeneous Enterprise Data Sources
Enterprise RAG & Vector Search 2026
2026-06-15
[ "cs.IR", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Enterprise business intelligence queries span structured warehouses and unstructured document repositories -- modalities with fundamentally different access methods, cost profiles, and correctness semantics. Existing AI-enabled interfaces force users
Darshita Rathore
5
[ "Darshita Rathore", "Vineet Kumar", "Vaibhav Singal", "Ankur Vivek Singh", "Anindya Moitra" ]
[]
http://arxiv.org/abs/2606.28370v1
VERIFIED_LIVE
https://github.com/unslothai/unsloth
[ "https://github.com/VectifyAI/PageIndex", "https://github.com/unslothai/unsloth" ]
70,823
0
Unknown
Unspecified
168.3
Enterprise business intelligence queries span structured warehouses and unstructured document repositories -- modalities with fundamentally different access methods, cost profiles, and correctness semantics. Existing AI-enabled interfaces force users to select the right tool: NL2SQL systems cannot reason over slide dec...
[ -0.006630000192672014, -0.023948999121785164, 0.07138899713754654, 0.0745140016078949, -0.037647999823093414, -0.09105999767780304, 0.05137600004673004, 0.0021069999784231186, -0.00020300000323913991, -0.06140900030732155, -0.016363000497221947, -0.05775599926710129, -0.0331059992313385, -...
[ -0.00964299961924553, -0.002432999899610877, 0.0033499998971819878, 0.05427800118923187, -0.057489000260829926, -0.047561001032590866, -0.013238999992609024, -0.0036239998880773783, 0.028473999351263046, 0.01723100058734417, -0.07371799647808075, -0.03918800130486488, 0.007083000149577856, ...
Enterprise RAG & Vector Search 2026
git clone https://github.com/unslothai/unsloth
Enterprise business intelligence queries span structured warehouses and unstructured document repositories -- modalities with fundamentally different access methods, cost profiles, and correctness semantics.
Existing AI-enabled interfaces force users to select the right tool: NL2SQL systems cannot reason over slide decks, and RAG pipelines lack access to live warehouse tables.
Finally, a caching layer validates query equivalence across multiple dimensions beyond embedding similarity, achieving zero false cache hits and $8.4\times$ latency reduction.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:17.235554
2607.01276v1
Embedding Inference Attack
Enterprise RAG & Vector Search 2026
2026-07-01
[ "cs.CR", "cs.IR", "cs.LG" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Embedding models are essential components of modern Information Retrieval (IR) systems, yet they are typically hidden behind APIs. Recent works have shown that dense IR system can lead to security vulnerabilities such as embedding inversion attacks.
Cedric Fitiavana Raelijohn
3
[ "Cedric Fitiavana Raelijohn", "Sébastien Gambs", "Jean-Francois Rajotte" ]
[]
http://arxiv.org/abs/2607.01276v1
VERIFIED_LIVE
https://github.com/Mintplex-Labs/anything-llm
[ "https://github.com/Mintplex-Labs/anything-llm", "https://github.com/Mintplex-Labs/anything-llm" ]
64,682
7,124
2026-08-13
MIT
168.12
Embedding models are essential components of modern Information Retrieval (IR) systems, yet they are typically hidden behind APIs. Recent works have shown that dense IR system can lead to security vulnerabilities such as embedding inversion attacks. However, such attacks usually require that the attacker knows the embe...
[ -0.051895998418331146, -0.024126999080181122, 0.05751200020313263, 0.008972999639809132, 0.12162099778652191, 0.09170500189065933, 0.07077900320291519, -0.04351799935102463, 0.024153999984264374, -0.001306000049225986, 0.04333199933171272, -0.04266899824142456, 0.1035429984331131, 0.016388...
[ -0.09789899736642838, -0.04328100010752678, -0.02076300047338009, 0.0585470013320446, 0.02658800035715103, 0.05767200142145157, -0.0014639999717473984, -0.023000000044703484, 0.08772800117731094, -0.06967899948358536, -0.007155000232160091, 0.01651199907064438, 0.1620739996433258, 0.002812...
Enterprise RAG & Vector Search 2026
git clone https://github.com/Mintplex-Labs/anything-llm
Embedding models are essential components of modern Information Retrieval (IR) systems, yet they are typically hidden behind APIs.
Recent works have shown that dense IR system can lead to security vulnerabilities such as embedding inversion attacks.
Finally, we propose and evaluate other mitigation strategies such as similarity thresholds.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:02.357258
2605.11188v1
Adversarial SQL Injection Generation with LLM-Based Architectures
Enterprise RAG & Vector Search 2026
2026-05-11
[ "cs.CR", "cs.AI", "cs.ET" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
SQL injection (SQLi) attacks are still one of the serious attacks ranked in the Open Worldwide Application Security Project (OWASP) Top 10 threats. Today, with advances in Artificial Intelligence (AI), especially in Large Language Models (LLMs), an o
Ali Karakoc
2
[ "Ali Karakoc", "H. Birkan Yilmaz" ]
[]
http://arxiv.org/abs/2605.11188v1
VERIFIED_LIVE
https://github.com/swisskyrepo/PayloadsAllTheThings
[ "https://github.com/BBVA/waf-brain", "https://github.com/swisskyrepo/PayloadsAllTheThings" ]
80,004
0
Unknown
Unspecified
167.88
SQL injection (SQLi) attacks are still one of the serious attacks ranked in the Open Worldwide Application Security Project (OWASP) Top 10 threats. Today, with advances in Artificial Intelligence (AI), especially in Large Language Models (LLMs), an opportunity has been created for automating adversarial attack tests to...
[ -0.03237900137901306, -0.08329000324010849, -0.0054720002226531506, 0.0659400001168251, -0.034237999469041824, -0.01041099987924099, 0.009344000369310379, -0.06768900156021118, -0.08200900256633759, -0.05297200009226799, 0.012790000066161156, -0.06978199630975723, 0.12287800014019012, -0.0...
[ -0.08316099643707275, -0.06067200005054474, -0.010967999696731567, 0.035211000591516495, 0.0348300002515316, -0.0682860016822815, 0.02329999953508377, -0.05821799859404564, -0.038155000656843185, -0.015204999595880508, -0.0328029990196228, -0.07676199823617935, 0.10936400294303894, -0.0353...
Enterprise RAG & Vector Search 2026
git clone https://github.com/swisskyrepo/PayloadsAllTheThings
SQL injection (SQLi) attacks are still one of the serious attacks ranked in the Open Worldwide Application Security Project (OWASP) Top 10 threats.
Today, with advances in Artificial Intelligence (AI), especially in Large Language Models (LLMs), an opportunity has been created for automating adversarial attack tests to measure the defense mechanisms.
In addition to these findings, another observation is that creating less diverse payloads achieves more bypasses, however they show poor results if the initially chosen payload is not successful.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:51.054738
2606.01697v2
RCEM: Robust Conversational Search EMbedder in Distributional Shift
Enterprise RAG & Vector Search 2026
2026-06-01
[ "cs.CL" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
We propose RCEM, a Robust Conversational search EMbedder that is additionally equipped with LLM's query reformulation capability without losing base model's generalization. Unlike prior conversational dense retrieval approaches that learn direct conv
Kilho Son
4
[ "Kilho Son", "Paul Hsu", "Cha Zhang", "Dinei Florencio" ]
[]
http://arxiv.org/abs/2606.01697v2
VERIFIED_LIVE
https://github.com/unslothai/unsloth
[ "https://github.com/unslothai/unsloth" ]
70,823
0
Unknown
Unspecified
167.6
We propose RCEM, a Robust Conversational search EMbedder that is additionally equipped with LLM's query reformulation capability without losing base model's generalization. Unlike prior conversational dense retrieval approaches that learn direct conversation-to-passage matching, RCEM aligns conversations, prepended by ...
[ -0.07436999678611755, -0.1192459985613823, 0.08159399777650833, 0.017645999789237976, 0.07596100121736526, 0.047586001455783844, -0.015069999732077122, 0.0012570000253617764, 0.08983399718999863, -0.06989700347185135, -0.06314600259065628, -0.07959699630737305, 0.0011390000581741333, 0.028...
[ -0.048211999237537384, -0.07494500279426575, 0.06872200220823288, 0.06477999687194824, 0.03168800100684166, 0.06362400203943253, -0.013914999552071095, 0.015506000258028507, 0.052365999668836594, -0.06896699965000153, -0.09592899680137634, -0.0030479999259114265, 0.01822800002992153, 0.037...
Enterprise RAG & Vector Search 2026
git clone https://github.com/unslothai/unsloth
We propose RCEM, a Robust Conversational search EMbedder that is additionally equipped with LLM's query reformulation capability without losing base model's generalization.
Unlike prior conversational dense retrieval approaches that learn direct conversation-to-passage matching, RCEM aligns conversations, prepended by special token, to LLM-rewritten queries, while preserving the original embedding space.
Extensive experiments show that RCEM consistently outperforms prior approaches, achieving up to 30% improvement under distributional shift.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:27.412219
2606.01041v1
ExpWeaver: LLM Agents Learn from Experience via Latent RAG
Enterprise RAG & Vector Search 2026
2026-05-31
[ "cs.CL" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Experience learning has achieved promising results in enhancing LLM agent planning and reasoning by integrating past interactions as reusable knowledge. However, existing methods remain confined to explicit text space, retrieving experiences via sema
Tao Feng
8
[ "Tao Feng", "Tianyang Luo", "Jingjun Xu", "Zhigang Hua", "Yan Xie", "Shuang Yang", "Ge Liu", "Jiaxuan You" ]
[]
http://arxiv.org/abs/2606.01041v1
VERIFIED_LIVE
https://github.com/unslothai/unsloth
[ "https://github.com/huggingface/trl", "https://github.com/unslothai/unsloth", "https://github.com/ulab-uiuc/ExpWeaver" ]
70,823
0
Unknown
Unspecified
167.55
Experience learning has achieved promising results in enhancing LLM agent planning and reasoning by integrating past interactions as reusable knowledge. However, existing methods remain confined to explicit text space, retrieving experiences via semantic similarity and concatenating them into the context window, leadin...
[ -0.02094700001180172, -0.008503000251948833, -0.022268999367952347, -0.01406800001859665, 0.023514999076724052, 0.009943000040948391, 0.03936599940061569, -0.04268600046634674, -0.012702000327408314, -0.009650999680161476, 0.018303999677300453, -0.030518999323248863, 0.04789299890398979, 0...
[ -0.03994299843907356, -0.0705069974064827, -0.006943999789655209, 0.08381500095129013, 0.039138998836278915, 0.007565999869257212, -0.002732000080868602, 0.01875999942421913, -0.0013500000350177288, -0.04048499837517738, -0.018059000372886658, -0.05203000083565712, 0.049897000193595886, 0....
Enterprise RAG & Vector Search 2026
git clone https://github.com/unslothai/unsloth
Experience learning has achieved promising results in enhancing LLM agent planning and reasoning by integrating past interactions as reusable knowledge.
However, existing methods remain confined to explicit text space, retrieving experiences via semantic similarity and concatenating them into the context window, leading to substantial token overhead and a decoupled architecture that separates retrieval from generation.
Results demonstrate that ExpWeaver achieves state-of-the-art performance on 12 out of 13 tasks, outperforming the strongest baseline by over 6.8%; maintains token efficiency comparable to non-retrieval baselines while text-based retrieval methods require 1.5 to 2 times more tokens; and exhibits superior cross-domain ge...
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:31.174959
2606.20047v1
PACMS: Submodular Context Selection as a Pluggable Engine for LLM Agents
Enterprise RAG & Vector Search 2026
2026-06-18
[ "cs.IR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Conversational and tool-using LLM agents operate over a context window that fills from several directions simultaneously. As a session proceeds, the agent accumulates user and assistant turns, entries drawn from a persistent memory store, and often l
Manu Ghulyani
5
[ "Manu Ghulyani", "Arunabh Singh", "Karan Bharadwaj", "Ankit Nath", "Suranjan Goswami" ]
[]
http://arxiv.org/abs/2606.20047v1
VERIFIED_LIVE
https://github.com/mem0ai/mem0
[ "https://github.com/run-llama/llama_index", "https://github.com/mem0ai/mem0", "https://github.com/langchain-ai/langchain" ]
63,191
0
Unknown
Unspecified
167.22
Conversational and tool-using LLM agents operate over a context window that fills from several directions simultaneously. As a session proceeds, the agent accumulates user and assistant turns, entries drawn from a persistent memory store, and often largest of all, the verbatim outputs of tool calls such as file reads, ...
[ 0.023249000310897827, -0.06288500130176544, -0.005849999841302633, -0.030981000512838364, 0.11794000118970871, -0.020208999514579773, 0.053266000002622604, 0.002566999988630414, -0.019447999075055122, -0.05335399881005287, -0.06657300144433975, -0.1033179983496666, 0.01686199940741062, 0.0...
[ -0.0015450000064447522, -0.026058999821543694, -0.05230399966239929, 0.044266000390052795, -0.0011180000146850944, 0.021331999450922012, 0.09570299834012985, 0.01358100026845932, 0.09084399789571762, -0.015616999939084053, 0.000012000000424450263, 0.056078001856803894, 0.05816100165247917, ...
Enterprise RAG & Vector Search 2026
git clone https://github.com/mem0ai/mem0
Conversational and tool-using LLM agents operate over a context window that fills from several directions simultaneously.
Once the cumulative context exceeds the model's token budget, the framework must decide what to keep.
Context-compression methods reduce token count by rewriting or pruning text, but operate query-blind and lossily.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:15.811928
2604.18946v1
Reasoning Structure Matters for Safety Alignment of Reasoning Models
AI Security, Red Teaming & Defense 2026
2026-04-21
[ "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Large reasoning models (LRMs) achieve strong performance on complex reasoning tasks but often generate harmful responses to malicious user queries. This paper investigates the underlying cause of these safety risks and shows that the issue lies in th
Yeonjun In
4
[ "Yeonjun In", "Wonjoong Kim", "Sangwu Park", "Chanyoung Park" ]
[]
http://arxiv.org/abs/2604.18946v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/yeonjun-in/R1-Act", "https://github.com/thu-coai/LRM-Safety-Study", "https://github.com/unslothai/unsloth" ]
70,809
0
Unknown
Unspecified
165.55
Large reasoning models (LRMs) achieve strong performance on complex reasoning tasks but often generate harmful responses to malicious user queries. This paper investigates the underlying cause of these safety risks and shows that the issue lies in the reasoning structure itself. Based on this insight, we claim that eff...
[ -0.012159000150859356, -0.04556399956345558, -0.06458699703216553, -0.05572099983692169, 0.07785999774932861, -0.006525000091642141, 0.04884500056505203, 0.10316500067710876, 0.017326999455690384, 0.05982600152492523, 0.03810499981045723, -0.03890499845147133, 0.08309199661016464, 0.059652...
[ -0.027726000174880028, -0.04217299818992615, -0.011246999725699425, -0.02123500034213066, 0.08709599822759628, 0.019050000235438347, 0.04466800019145012, 0.06905300170183182, 0.0009549999958835542, 0.06293000280857086, -0.016445999965071678, -0.048117998987436295, 0.11209700256586075, 0.04...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Large reasoning models (LRMs) achieve strong performance on complex reasoning tasks but often generate harmful responses to malicious user queries.
We propose AltTrain, a simple yet effective post training method that explicitly alters the reasoning structure of LRMs.
Experiments across LRM backbones and model sizes demonstrate strong safety alignment, along with robust generalization across reasoning, QA, summarization, and multilingual setting.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:46:24.859168
2606.21071v1
Local LLM Agents as Vulnerable Runtimes:A Source-Code Audit of the Agent Runtime Layer
AI Security, Red Teaming & Defense 2026
2026-06-19
[ "cs.CR", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Local LLM agents such as OpenClaw and Nanobot run on end-user machines and act on host resources - the shell, filesystem, browser, stored credentials, and messaging applications - through natural-language goals. These agents have become privileged so
Zhengsong Zhang
4
[ "Zhengsong Zhang", "Zongze Li", "Jiawei Guo", "Haipeng Cai" ]
[]
http://arxiv.org/abs/2606.21071v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/snyk/agent-scan", "https://github.com/jerryjliu/llama_index", "https://github.com/HKUDS/nanobot" ]
51,616
0
Unknown
Unspecified
165.07
Local LLM agents such as OpenClaw and Nanobot run on end-user machines and act on host resources - the shell, filesystem, browser, stored credentials, and messaging applications - through natural-language goals. These agents have become privileged software runtimes that mediate between user intent, model outputs, and h...
[ -0.033309999853372574, 0.0002809999859891832, -0.04756699874997139, -0.0018149999668821692, 0.05445199832320213, -0.011486000381410122, -0.012116000056266785, -0.027566999197006226, -0.012198000214993954, -0.02521499991416931, 0.025561999529600143, -0.03651599958539009, 0.02600800059735775, ...
[ 0.00021699999342672527, -0.007286999840289354, -0.055984001606702805, -0.0025059999898076057, 0.06245400011539459, -0.04848499968647957, -0.042879000306129456, 0.016473999246954918, -0.04089599847793579, -0.007240999955683947, -0.014356999658048153, -0.04932200163602829, 0.07829000055789948,...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Local LLM agents such as OpenClaw and Nanobot run on end-user machines and act on host resources - the shell, filesystem, browser, stored credentials, and messaging applications - through natural-language goals.
These agents have become privileged software runtimes that mediate between user intent, model outputs, and host-level actions.
We instantiate the taxonomy in two backends, 47 Semgrep YAML rules and 30 CodeQL queries, and evaluate on OPENCLAWBENCH, a benchmark of 446 source-code-level advisories from the OpenClaw repository and split temporally into 229 rule-derivation (train) and 217 held-out (test) advisories.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:24:53.423420
2606.00101v1
CoCoVideo: The High-Quality Commercial-Model-Based Contrastive Benchmark for AI-Generated Video Detection
AI Security, Red Teaming & Defense 2026
2026-05-26
[ "cs.CV", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
With the rapid advancement of artificial intelligence generated content (AIGC) technologies, video forgery has become increasingly prevalent, posing new challenges to public discourse and societal security. Despite remarkable progress in existing dee
Huidong Feng
8
[ "Huidong Feng", "Wentao Chen", "Jie Chen", "Xinqi Cai", "Ruolong Ma", "Yinglin Zheng", "Yuxin Lin", "Ming Zeng" ]
[]
http://arxiv.org/abs/2606.00101v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/DonoToT/CoCoVideo", "https://github.com/MarekKowalski/FaceSwap", "https://github.com/deepfakes/faceswap" ]
57,447
0
Unknown
Unspecified
165.03
With the rapid advancement of artificial intelligence generated content (AIGC) technologies, video forgery has become increasingly prevalent, posing new challenges to public discourse and societal security. Despite remarkable progress in existing deepfake detection methods, AIGC forgery detection remains challenging, a...
[ 0.051148999482393265, -0.14800900220870972, -0.01460499968379736, 0.040835000574588776, 0.13649700582027435, 0.028505999594926834, -0.04225799813866615, -0.04166200011968613, 0.08085999637842178, -0.0835219994187355, -0.06248300150036812, -0.11841800063848495, -0.029650000855326653, 0.0677...
[ -0.05804400146007538, -0.12974999845027924, 0.013299999758601189, 0.028279999271035194, 0.11948399990797043, -0.012481000274419785, -0.04013799875974655, -0.08847499638795853, 0.09188800305128098, -0.09317599982023239, -0.030775999650359154, -0.06946399807929993, -0.00873200036585331, 0.05...
AI Security, Red Teaming & Defense 2026
No code repository attached.
With the rapid advancement of artificial intelligence generated content (AIGC) technologies, video forgery has become increasingly prevalent, posing new challenges to public discourse and societal security.
Despite remarkable progress in existing deepfake detection methods, AIGC forgery detection remains challenging, as existing datasets mainly rely on open-source video generation models with quality far below that of commercial AIGC systems.
Extensive experiments on CoCoVideo-26K and public benchmarks demonstrate state-of-the-art performance, validating the framework's robustness and generalizability.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:33:17.560026
2605.03482v2
MEMSAD: Gradient-Coupled Anomaly Detection for Memory Poisoning in Retrieval-Augmented Agents
Enterprise RAG & Vector Search 2026
2026-05-05
[ "cs.CR", "cs.AI", "cs.LG" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Persistent external memory enables LLM agents to maintain context across sessions, yet its security properties remain formally uncharacterized. We formalize memory poisoning attacks on retrieval-augmented agents as a Stackelberg game with a unified e
Ishrith Gowda
1
[ "Ishrith Gowda" ]
[ "University of California, Berkeley" ]
http://arxiv.org/abs/2605.03482v2
VERIFIED_LIVE
https://github.com/mem0ai/mem0
[ "https://github.com/agiresearch/A-MEM", "https://github.com/mem0ai/mem0" ]
63,191
0
Unknown
Unspecified
165.02
Persistent external memory enables LLM agents to maintain context across sessions, yet its security properties remain formally uncharacterized. We formalize memory poisoning attacks on retrieval-augmented agents as a Stackelberg game with a unified evaluation framework spanning three attack classes with escalating acce...
[ -0.006597999949008226, -0.07800400257110596, -0.014756999909877777, 0.03165699914097786, 0.05350600183010101, -0.011374999769032001, 0.10469699651002884, -0.002094999887049198, 0.022211000323295593, 0.013620000332593918, 0.04377000033855438, 0.011611999943852425, 0.09374099969863892, 0.046...
[ -0.049320001155138016, -0.04636500030755997, -0.07787200063467026, 0.011251999996602535, 0.07048399746417999, 0.018363000825047493, 0.06598400324583054, 0.00022600000374950469, 0.04843999817967415, 0.038040999323129654, -0.012401999905705452, -0.04377799853682518, 0.1230119988322258, 0.011...
Enterprise RAG & Vector Search 2026
git clone https://github.com/mem0ai/mem0
Persistent external memory enables LLM agents to maintain context across sessions, yet its security properties remain formally uncharacterized.
We formalize memory poisoning attacks on retrieval-augmented agents as a Stackelberg game with a unified evaluation framework spanning three attack classes with escalating access assumptions.
Experiments on a $3 \times 5$ attack-defense matrix with bootstrap confidence intervals, Bonferroni-corrected hypothesis tests, and Clopper-Pearson validation ($n=1{,}000$) confirm: composite defenses achieve TPR $= 1.00$, FPR $= 0.00$ across all attacks, while synonym substitution evades detection at $Δ$ ASR-R $\appro...
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:57.350924
2605.19743v2
EngiAI: A Multi-Agent Framework and Benchmark Suite for LLM-Driven Engineering Design
Enterprise RAG & Vector Search 2026
2026-05-19
[ "cs.AI", "cs.LG", "cs.MA" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Large Language Model (LLM) agents are increasingly applied to engineering design tasks, yet existing evaluation frameworks do not adequately address multi-agent systems that combine simulation, retrieval, and manufacturing preparation. We introduce a
Gioele Molinari
4
[ "Gioele Molinari", "Florian Felten", "Soheyl Massoudi", "Mark Fuge" ]
[]
http://arxiv.org/abs/2605.19743v2
VERIFIED_LIVE
https://github.com/crewAIInc/crewAI
[ "https://github.com/langchain-ai/langgraph", "https://github.com/crewAIInc/crewAI", "https://github.com/langchain-ai/langchain", "https://github.com/openai/openai-agents-python" ]
57,030
0
Unknown
Unspecified
164.6
Large Language Model (LLM) agents are increasingly applied to engineering design tasks, yet existing evaluation frameworks do not adequately address multi-agent systems that combine simulation, retrieval, and manufacturing preparation. We introduce a benchmark suite with three evaluation dimensions: (1) a workflow benc...
[ 0.026623999699950218, -0.011211000382900238, -0.03030399978160858, -0.026142999529838562, 0.020840000361204147, -0.08799900114536285, -0.0672840029001236, -0.03057599999010563, -0.01966400071978569, -0.056919001042842865, -0.07780499756336212, -0.09095600247383118, 0.058348000049591064, 0....
[ 0.0053079999051988125, -0.0616380013525486, 0.0001720000000204891, 0.015185999684035778, -0.03616299852728844, -0.04508399963378906, -0.06102199852466583, 0.001057000015862286, -0.01447299961000681, -0.07819300144910812, -0.06305699795484543, -0.023090999573469162, 0.013673000037670135, 0....
Enterprise RAG & Vector Search 2026
git clone https://github.com/crewAIInc/crewAI
Large Language Model (LLM) agents are increasingly applied to engineering design tasks, yet existing evaluation frameworks do not adequately address multi-agent systems that combine simulation, retrieval, and manufacturing preparation.
We introduce a benchmark suite with three evaluation dimensions: (1) a workflow benchmark with seven prompt styles targeting distinct cognitive demands-including direct tool use, semantic disambiguation, conditional branching, and working-memory tasks; (2) a Retrieval-Augmented Generation (RAG) benchmark with gated sco...
Across four LLM backends and two EngiBench problems, proprietary models achieve 96-97% average task completion on Beams2D, while open-source 4B-parameter models reach 55-78%, with clear generational improvement.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:45.387793
2606.28002v1
Dialogue to Detection: A Multimodal Hybrid NLP Pipeline for Insurance Fraud Detection
Enterprise RAG & Vector Search 2026
2026-06-26
[ "cs.CL", "cs.AI", "eess.AS" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Insurance fraud imposes substantial financial losses and operational inefficiencies, raising premiums and impacting trust among legitimate policyholders. Early detection at FNOL remains a persistent challenge. Existing approaches rely largely on priv
Muhammad Shakeel Akram
5
[ "Muhammad Shakeel Akram", "Amal Htait", "Abdul Hamid Sadka", "Emma Meisingseth", "Karishma Jaitly" ]
[]
http://arxiv.org/abs/2606.28002v1
VERIFIED_LIVE
https://github.com/coqui-ai/TTS
[ "https://github.com/PrajwalSuryaPrakash/Insurance-Claims-Fraud-Detection-Model", "https://github.com/coqui-ai/TTS", "https://github.com/resemble-ai/Resemblyzer", "https://github.com/yingcheuk/fraud-detector" ]
45,890
0
Unknown
Unspecified
164.14
Insurance fraud imposes substantial financial losses and operational inefficiencies, raising premiums and impacting trust among legitimate policyholders. Early detection at FNOL remains a persistent challenge. Existing approaches rely largely on private, text-only datasets, limiting progress on multimodal methods that ...
[ -0.04646499827504158, 0.03279099985957146, 0.007455000188201666, -0.048505999147892, 0.09038600325584412, -0.0008559999987483025, 0.10504300147294998, 0.0062060002237558365, 0.07363499701023102, -0.034449998289346695, -0.001601999974809587, -0.06344500184059143, 0.01933399960398674, 0.0322...
[ -0.031661998480558395, -0.05853300169110298, -0.010201999917626381, 0.015227999538183212, 0.11218100041151047, 0.0555340014398098, 0.09069299697875977, -0.04794999957084656, 0.10262099653482437, -0.07899399846792221, 0.005167999770492315, -0.04011499881744385, 0.02280599996447563, 0.037307...
Enterprise RAG & Vector Search 2026
git clone https://github.com/coqui-ai/TTS
Insurance fraud imposes substantial financial losses and operational inefficiencies, raising premiums and impacting trust among legitimate policyholders.
Existing approaches rely largely on private, text-only datasets, limiting progress on multimodal methods that integrate linguistic, behavioural, and speaker-based indicators.
Dataset validation and component-level evaluations show stability and transfer potential, offering a reproducible baseline beyond text-only fraud detection.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:13.463162
2605.05287v1
Securing the Agent: Vendor-Neutral, Multitenant Enterprise Retrieval and Tool Use
Enterprise RAG & Vector Search 2026
2026-05-06
[ "cs.CR", "cs.AI", "cs.IR", "cs.SE" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Retrieval-Augmented Generation (RAG) and agentic AI systems are increasingly prevalent in enterprise AI deployments. However, real enterprise environments introduce challenges largely absent from academic treatments and consumer-facing APIs: multiple
Francisco Javier Arceo
2
[ "Francisco Javier Arceo", "Varsha Prasad Narsing" ]
[ "OpenAI" ]
http://arxiv.org/abs/2605.05287v1
VERIFIED_LIVE
https://github.com/crewAIInc/crewAI
[ "https://github.com/crewAIInc/crewAI", "https://github.com/microsoft/autogen", "https://github.com/microsoft/semantic-kernel", "https://github.com/pydantic/pydantic-ai" ]
57,030
0
Unknown
Unspecified
163.95
Retrieval-Augmented Generation (RAG) and agentic AI systems are increasingly prevalent in enterprise AI deployments. However, real enterprise environments introduce challenges largely absent from academic treatments and consumer-facing APIs: multiple tenants with heterogeneous data, strict access-control requirements, ...
[ -0.07657399773597717, 0.05162699893116951, -0.08785899728536606, -0.06396400183439255, -0.0066019999794662, -0.0026110000908374786, 0.08052200078964233, -0.0051830001175403595, 0.041103001683950424, -0.05000400170683861, 0.008617999963462353, -0.01794400066137314, 0.06506100296974182, -0.0...
[ -0.06598600000143051, 0.02663400024175644, -0.027083000168204308, 0.040084000676870346, -0.007278000004589558, -0.027056999504566193, 0.006574000231921673, -0.0247189998626709, 0.06981900334358215, -0.04002700001001358, -0.016015000641345978, -0.03294700011610985, 0.06171900033950806, -0.0...
Enterprise RAG & Vector Search 2026
git clone https://github.com/crewAIInc/crewAI
Retrieval-Augmented Generation (RAG) and agentic AI systems are increasingly prevalent in enterprise AI deployments.
However, real enterprise environments introduce challenges largely absent from academic treatments and consumer-facing APIs: multiple tenants with heterogeneous data, strict access-control requirements, regulatory compliance, and cost pressures that demand shared infrastructure.
We evaluate it empirically and show that ABAC gating eliminates cross-tenant leakage while introducing negligible overhead.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:58.624033
2605.27220v1
The Coverage Illusion: From Pre-retrieval Routing Failure to Post-retrieval Cascades in a Production RAG System
Enterprise RAG & Vector Search 2026
2026-05-26
[ "cs.CL", "cs.IR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
In modern RAG pipelines, query augmentation methods such as HyDE and query expansion are applied to every query, resulting in substantial LLM inference costs and increased end-to-end latency. The empirical justification for this overhead in real prod
Zafar Hussain
2
[ "Zafar Hussain", "Kristoffer Nielbo" ]
[]
http://arxiv.org/abs/2605.27220v1
VERIFIED_LIVE
https://github.com/jerryjliu/llama_index
[ "https://github.com/jerryjliu/llama_index", "https://github.com/langchain-ai/langchain" ]
51,617
0
Unknown
Unspecified
163.87
In modern RAG pipelines, query augmentation methods such as HyDE and query expansion are applied to every query, resulting in substantial LLM inference costs and increased end-to-end latency. The empirical justification for this overhead in real production traffic remains largely unexplored. We present a case study of ...
[ -0.023545999079942703, 0.041912999004125595, 0.07340600341558456, -0.0056929998099803925, -0.05936700105667114, -0.006732999812811613, -0.02878599986433983, -0.04288500174880028, -0.04763900116086006, -0.0034630000591278076, 0.012412000447511673, 0.06773500144481659, 0.0704369992017746, -0...
[ -0.06474799662828445, -0.04539699852466583, 0.03961100056767464, 0.04451100155711174, -0.03186599910259247, -0.01682800054550171, -0.017215000465512276, -0.024481000378727913, -0.020222999155521393, -0.04570300132036209, -0.06282400339841843, 0.020006999373435974, 0.018849000334739685, -0....
Enterprise RAG & Vector Search 2026
git clone https://github.com/jerryjliu/llama_index
In modern RAG pipelines, query augmentation methods such as HyDE and query expansion are applied to every query, resulting in substantial LLM inference costs and increased end-to-end latency.
We present a case study of the Danish National Encyclopedia, evaluating five retrieval workflows over 20,000 query-workflow pairs from production traffic and synthetic conditions.
Operating entirely without training overhead or secondary serving infrastructure, the cascade improves quality by +0.140 Composite Overall points over Always-HyDE, reduces latency by 31.8%, and serves 72.2% of real user queries without LLM augmentation.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:35.725923
2605.01186v1
Trace: Unmasking AI Attack Agents Through Terminal Behavior Fingerprinting
AI Security, Red Teaming & Defense 2026
2026-05-02
[ "cs.CR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
AI-driven penetration testing agents are now capable of autonomously executing attacks within compromised networks. Identifying the model family that controls the active sessions of such agents provides valuable information towards understanding the
Murali Ediga
2
[ "Murali Ediga", "Sudipta Chattopadhyay" ]
[]
http://arxiv.org/abs/2605.01186v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/BerriAI/litellm" ]
56,255
0
Unknown
Unspecified
163.6
AI-driven penetration testing agents are now capable of autonomously executing attacks within compromised networks. Identifying the model family that controls the active sessions of such agents provides valuable information towards understanding the intent of the attack and further developing attack countermeasures. In...
[ -0.04105899855494499, -0.0417260006070137, -0.04971899837255478, -0.010193999856710434, 0.07917000353336334, -0.00684700021520257, 0.0527070015668869, -0.05934999883174896, -0.0224239993840456, -0.022437000647187233, -0.004081999883055687, -0.062387000769376755, 0.02506300061941147, 0.0171...
[ -0.06194499880075455, -0.014433000236749649, -0.06128200143575668, -0.0039470000192523, 0.10276100039482117, 0.002764000091701746, 0.02258799970149994, -0.009309000335633755, 0.05069499835371971, 0.022355999797582626, 0.016812000423669815, -0.06822899729013443, 0.06134700030088425, 0.01047...
AI Security, Red Teaming & Defense 2026
No code repository attached.
AI-driven penetration testing agents are now capable of autonomously executing attacks within compromised networks.
Identifying the model family that controls the active sessions of such agents provides valuable information towards understanding the intent of the attack and further developing attack countermeasures.
Finally, to validate the robustness of Trace, we evaluate it with a blackbox and proprietary scaffold employing multiple model families (Gemini and Claude Opus).
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:43:03.865888
2605.05244v1
Towards Dependable Retrieval-Augmented Generation Using Factual Confidence Prediction
Enterprise RAG & Vector Search 2026
2026-05-04
[ "cs.IR", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Incorporating specific knowledge into large language models via retrieval-augmented generation (RAG) is a widespread technique that fuels many of today's industry AI applications. A fundamental problem is to assess if the context retrieved by some si
Florian Geissler
4
[ "Florian Geissler", "Francesco Carella", "Laura Fieback", "Jakob Spiegelberg" ]
[]
http://arxiv.org/abs/2605.05244v1
VERIFIED_LIVE
https://github.com/jerryjliu/llama_index
[ "https://github.com/jerryjliu/llama_index", "https://github.com/lauhaide/ragu" ]
51,617
0
Unknown
Unspecified
162.77
Incorporating specific knowledge into large language models via retrieval-augmented generation (RAG) is a widespread technique that fuels many of today's industry AI applications. A fundamental problem is to assess if the context retrieved by some similarity search provides indeed supporting facts, or instead misguides...
[ -0.09683900326490402, -0.01395499985665083, -0.037735000252723694, 0.09568099677562714, 0.030229000374674797, 0.08661600202322006, 0.03614800050854683, 0.020788000896573067, -0.0028899998869746923, -0.019514000043272972, 0.03265700116753578, -0.08894599974155426, 0.0845789983868599, -0.037...
[ -0.09171699732542038, -0.05499500036239624, -0.04310699924826622, 0.07491400092840195, 0.046535998582839966, 0.033222001045942307, 0.051711998879909515, 0.014179999940097332, 0.009821999818086624, -0.051878999918699265, -0.006821999792009592, -0.0687979981303215, 0.07302899658679962, -0.05...
Enterprise RAG & Vector Search 2026
git clone https://github.com/jerryjliu/llama_index
Incorporating specific knowledge into large language models via retrieval-augmented generation (RAG) is a widespread technique that fuels many of today's industry AI applications.
We present a new, two-staged approach to predict fact faithfulness of the output of retrieval-augmented generations.
This approach in itself can improve answer quality by up to 6% in some of the studied datasets, however, the associated statistical guarantees do not hold generally, since the assumption of sample exchangeability depends on the retriever setup.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:58.931133
2605.30686v1
Depth-Dependent Indirect Prompt Injection in Tool-Calling ReAct Agents: Injection Depth, Payload Framing, and Turn-Budget Sensitivity
AI Security, Red Teaming & Defense 2026
2026-05-29
[ "cs.CR", "cs.AI", "cs.LG" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
ReAct agents that interleave chain-of-thought reasoning with tool calls are increasingly deployed for real tasks such as scheduling, file retrieval, and data access. Their tool observation loop creates a direct attack surface: an adversary who contro
Mohammadreza Rashidi
1
[ "Mohammadreza Rashidi" ]
[]
http://arxiv.org/abs/2605.30686v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/langchain-ai/langgraph" ]
39,609
0
Unknown
Unspecified
161.15
ReAct agents that interleave chain-of-thought reasoning with tool calls are increasingly deployed for real tasks such as scheduling, file retrieval, and data access. Their tool observation loop creates a direct attack surface: an adversary who controls any tool's return value can embed instructions that redirect the ag...
[ -0.031394001096487045, -0.0032919999212026596, 0.007017000112682581, -0.0018390000332146883, -0.056974999606609344, -0.06430300325155258, -0.016589000821113586, 0.09364999830722809, 0.04243899881839752, -0.02999800071120262, -0.07348199933767319, -0.06633800268173218, 0.010041999630630016, ...
[ -0.07426100224256516, 0.0038799999747425318, -0.015943000093102455, 0.027215000241994858, 0.017795000225305557, -0.051697999238967896, -0.006984000094234943, 0.07999099791049957, 0.07590299844741821, 0.008965999819338322, -0.08369500190019608, -0.02308100089430809, 0.015022000297904015, 0....
AI Security, Red Teaming & Defense 2026
No code repository attached.
ReAct agents that interleave chain-of-thought reasoning with tool calls are increasingly deployed for real tasks such as scheduling, file retrieval, and data access.
Their tool observation loop creates a direct attack surface: an adversary who controls any tool's return value can embed instructions that redirect the agent away from the user's goal, a threat known as indirect prompt injection.
Our results establish injection depth as the dominant variable and show that sanitising only the first tool observation captures 67% of measured injection successes.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:31:30.755490
2605.29224v1
Relevance as a Vulnerability: How Web Retrieval Degrades Safety Alignment in LLM Agents
AI Security, Red Teaming & Defense 2026
2026-05-28
[ "cs.CL", "cs.AI", "cs.CR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
AI agents augment large language models with external tools such as web retrieval, enabling grounded and up-to-date responses. However, incorporating external content into the generation pipeline can weaken the safety alignment mechanisms that govern
Aditya Nawal
3
[ "Aditya Nawal", "Manit Baser", "Mohan Gurusamy" ]
[]
http://arxiv.org/abs/2605.29224v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/langchain-ai/langgraph" ]
39,609
0
Unknown
Unspecified
161.1
AI agents augment large language models with external tools such as web retrieval, enabling grounded and up-to-date responses. However, incorporating external content into the generation pipeline can weaken the safety alignment mechanisms that govern model outputs. Prior work shows that enabling retrieval in agents inc...
[ -0.024322999641299248, -0.012148999609053135, -0.09070699661970139, -0.03558100014925003, 0.06695999950170517, 0.025713000446558, 0.02409300021827221, 0.007687999866902828, 0.036775000393390656, -0.0550989992916584, 0.037769999355077744, 0.014278000220656395, 0.09569799900054932, -0.045471...
[ -0.07299300283193588, -0.054892998188734055, -0.0538100004196167, 0.004422999918460846, 0.07617899775505066, 0.03341599926352501, -0.05445199832320213, 0.014209000393748283, 0.04207700118422508, -0.05597100034356117, -0.02698799967765808, -0.021129999309778214, 0.10117900371551514, -0.0325...
AI Security, Red Teaming & Defense 2026
No code repository attached.
AI agents augment large language models with external tools such as web retrieval, enabling grounded and up-to-date responses.
However, incorporating external content into the generation pipeline can weaken the safety alignment mechanisms that govern model outputs.
Because relevance is also what makes retrieval useful, these results expose a safety-utility trade-off for retrieval-enabled agents.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:32:18.741069
2605.06285v1
LatentRAG: Latent Reasoning and Retrieval for Efficient Agentic RAG
Enterprise RAG & Vector Search 2026
2026-05-07
[ "cs.CL", "cs.LG" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Single-step retrieval-augmented generation (RAG) provides an efficient way to incorporate external information for simple question answering tasks but struggles with complex questions. Agentic RAG extends this paradigm by replacing single-step retrie
Yijia Zheng
2
[ "Yijia Zheng", "Marcel Worring" ]
[]
http://arxiv.org/abs/2605.06285v1
VERIFIED_LIVE
https://github.com/deepspeedai/DeepSpeed
[ "https://github.com/deepspeedai/DeepSpeed", "https://github.com/microsoft/LoRA", "https://github.com/dao-ailab/flash-attention", "https://github.com/facebookresearch/faiss" ]
42,925
0
Unknown
Unspecified
160.92
Single-step retrieval-augmented generation (RAG) provides an efficient way to incorporate external information for simple question answering tasks but struggles with complex questions. Agentic RAG extends this paradigm by replacing single-step retrieval with a multi-step process, in which the large language model (LLM)...
[ -0.01191100012511015, 0.01423299964517355, -0.01863499917089939, -0.04077399894595146, 0.032301001250743866, 0.05174599960446358, 0.036945998668670654, -0.017638999968767166, 0.040824998170137405, 0.019199000671505928, 0.0032099999953061342, -0.03808499872684479, 0.05506100133061409, 0.103...
[ -0.04522399976849556, -0.04588500037789345, 0.01055699959397316, 0.020270999521017075, -0.0022559999488294125, 0.03889999911189079, -0.015595999546349049, 0.022071000188589096, 0.07559599727392197, -0.033048998564481735, -0.0017089999746531248, -0.05834700167179108, 0.05527399852871895, -0...
Enterprise RAG & Vector Search 2026
git clone https://github.com/deepspeedai/DeepSpeed
Single-step retrieval-augmented generation (RAG) provides an efficient way to incorporate external information for simple question answering tasks but struggles with complex questions.
Agentic RAG extends this paradigm by replacing single-step retrieval with a multi-step process, in which the large language model (LLM) acts as a search agent that generates intermediate thoughts and subqueries to iteratively interact with the retrieval system.
Extensive experiments on seven benchmark datasets show that LatentRAG achieves performance comparable to explicit agentic RAG methods while reducing inference latency by approximately 90%, substantially narrowing the latency gap with traditional single-step RAG.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:24:01.504020
2605.30966v1
Reading Between the Citations: A Typed Claim Network for Scientific Literature
Enterprise RAG & Vector Search 2026
2026-05-29
[ "cs.IR", "cs.AI", "cs.CL" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Knowledge graphs over corpora of inter-referencing documents - scholarly papers, legal opinions, policy briefs - encode the topology of reference but not its stance. The standard representation collapses a rich evaluative relation into an untyped edg
Ning Ding
3
[ "Ning Ding", "Sergio J. Rodríguez Méndez", "Pouya G. Omran" ]
[]
http://arxiv.org/abs/2605.30966v1
VERIFIED_LIVE
https://github.com/datalab-to/marker
[ "https://github.com/datalab-to/marker" ]
38,712
0
Unknown
Unspecified
160.9
Knowledge graphs over corpora of inter-referencing documents - scholarly papers, legal opinions, policy briefs - encode the topology of reference but not its stance. The standard representation collapses a rich evaluative relation into an untyped edge, losing the very content that supports community-level queries about...
[ -0.0885080024600029, -0.040052998811006546, 0.00775599991902709, 0.04408299922943115, 0.08109600096940994, 0.01776999980211258, 0.025662999600172043, 0.09837699681520462, 0.08575599640607834, 0.09112899750471115, -0.025182999670505524, 0.014333999715745449, -0.0013930000131949782, 0.009267...
[ -0.07312300056219101, -0.03592799976468086, -0.04972200095653534, 0.027096999809145927, 0.060015998780727386, 0.02823599986732006, 0.03671199828386307, 0.0377580001950264, 0.06168299913406372, 0.05085200071334839, -0.05619800090789795, 0.03210299834609032, 0.03486299887299538, 0.0697960034...
Enterprise RAG & Vector Search 2026
git clone https://github.com/datalab-to/marker
Knowledge graphs over corpora of inter-referencing documents - scholarly papers, legal opinions, policy briefs - encode the topology of reference but not its stance.
The standard representation collapses a rich evaluative relation into an untyped edge, losing the very content that supports community-level queries about how one document is received by another.
Three downstream task families demonstrate what the network enables: retrieval signal augmentation, aggregated-stance summarisation, and topological analytics.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:32.815360
2605.18850v1
KadiAssistant: A conversational AI Agent for information retrieval in Kadi4Mat
Enterprise RAG & Vector Search 2026
2026-05-13
[ "cs.IR", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
We introduce KadiAssistant, a privacy-by-design AI assistant integrated into the Kadi research data ecosystem, enabling researchers to efficiently access, aggregate, and synthesize information from heterogeneous, privacy-sensitive research data. Inte
Adrian Cierpka
6
[ "Adrian Cierpka", "Mohammad Shafiqul Islam", "Johannes Steinhülb", "Eric Dietriche Sesso Domtchoueng", "Michael Selzer", "Arnd Koeppe" ]
[ "Meta AI (FAIR)" ]
http://arxiv.org/abs/2605.18850v1
VERIFIED_LIVE
https://github.com/langchain-ai/langgraph
[ "https://github.com/langchain-ai/langgraph", "https://github.com/pgvector/pgvector", "https://github.com/Unstructured-IO/unstructured" ]
39,611
0
Unknown
Unspecified
160.35
We introduce KadiAssistant, a privacy-by-design AI assistant integrated into the Kadi research data ecosystem, enabling researchers to efficiently access, aggregate, and synthesize information from heterogeneous, privacy-sensitive research data. Interdisciplinary fields such as materials science bring together discipli...
[ -0.04586799815297127, -0.0075240000151097775, -0.006938000209629536, 0.009217999875545502, -0.014138000085949898, -0.00597799988463521, 0.016388000920414925, 0.0060339998453855515, 0.02132900059223175, -0.04325199872255325, -0.019229000434279442, -0.09247799962759018, 0.002391000045463443, ...
[ -0.08934500068426132, 0.02149300090968609, -0.0694269984960556, 0.05699300020933151, 0.005963000003248453, -0.018893999978899956, -0.07597500085830688, 0.0001230000052601099, 0.022135000675916672, 0.029148999601602554, 0.01032199990004301, -0.042603999376297, 0.0780080035328865, 0.00212799...
Enterprise RAG & Vector Search 2026
git clone https://github.com/langchain-ai/langgraph
We introduce KadiAssistant, a privacy-by-design AI assistant integrated into the Kadi research data ecosystem, enabling researchers to efficiently access, aggregate, and synthesize information from heterogeneous, privacy-sensitive research data.
Interdisciplinary fields such as materials science bring together disciplines with their own terminology and standards.
KadiAssistant therefore bridges terminology and standards, lowers access barriers for researchers, and strengthens the Findable pillar of FAIR data principles.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:45.749381
2605.13295v1
CANTANTE: Optimizing Agentic Systems via Contrastive Credit Attribution
Enterprise RAG & Vector Search 2026
2026-05-13
[ "cs.CL", "cs.AI", "cs.MA" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
LLM-based multi-agent systems have demonstrated strong performance across complex real-world tasks, such as software engineering, predictive modeling, and retrieval-augmented generation. Yet automating their configuration remains a structural challen
Tom Zehle
1
[ "Tom Zehle" ]
[]
http://arxiv.org/abs/2605.13295v1
VERIFIED_LIVE
https://github.com/langchain-ai/langgraph
[ "https://github.com/langchain-ai/langgraph", "https://github.com/finitearth/cantante" ]
39,611
0
Unknown
Unspecified
160.35
LLM-based multi-agent systems have demonstrated strong performance across complex real-world tasks, such as software engineering, predictive modeling, and retrieval-augmented generation. Yet automating their configuration remains a structural challenge, as scores are available only at the system level, whereas the para...
[ -0.04239499941468239, 0.0073640001937747, -0.07915200293064117, -0.06813900172710419, 0.0884540006518364, 0.015070999972522259, 0.044353000819683075, -0.007493999786674976, 0.04928499832749367, 0.05238699913024902, -0.02930299937725067, -0.030293000862002373, 0.09731599688529968, 0.0593639...
[ -0.03530700132250786, -0.03951700031757355, -0.07741200178861618, -0.022505000233650208, 0.03288000077009201, 0.003544999985024333, 0.022551000118255615, -0.002598999999463558, -0.010600999929010868, 0.027202999219298363, -0.08844900131225586, -0.06129800155758858, 0.10735700279474258, 0.0...
Enterprise RAG & Vector Search 2026
git clone https://github.com/langchain-ai/langgraph
LLM-based multi-agent systems have demonstrated strong performance across complex real-world tasks, such as software engineering, predictive modeling, and retrieval-augmented generation.
Yet automating their configuration remains a structural challenge, as scores are available only at the system level, whereas the parameters governing agent behavior are local.
It improves over the strongest baseline by +18.9 percentage points on MBPP and +12.5 percentage points on GSM8K, while incurring a lower inference cost.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:47.558833
2605.09661v1
MedMeta: A Benchmark for LLMs in Synthesizing Meta-Analysis Conclusion from Medical Studies
Enterprise RAG & Vector Search 2026
2026-05-10
[ "cs.CL", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Large language models (LLMs) have saturated standard medical benchmarks that test factual recall, yet their ability to perform higher-order reasoning, such as synthesizing evidence from multiple sources, remains critically under-explored. To address
Huy Hoang Ha
3
[ "Huy Hoang Ha", "Benoit Favre", "Francois Portet" ]
[]
http://arxiv.org/abs/2605.09661v1
VERIFIED_LIVE
https://github.com/langchain-ai/langgraph
[ "https://github.com/langchain-ai/langgraph" ]
39,611
0
Unknown
Unspecified
160.2
Large language models (LLMs) have saturated standard medical benchmarks that test factual recall, yet their ability to perform higher-order reasoning, such as synthesizing evidence from multiple sources, remains critically under-explored. To address this gap, we introduce MedMeta, the first benchmark designed to evalua...
[ 0.08390100300312042, -0.01030800025910139, -0.024939000606536865, 0.005061999894678593, 0.0466499999165535, 0.0304189994931221, -0.07009100168943405, 0.1628270000219345, 0.03985600173473358, 0.034940000623464584, -0.06747200340032578, 0.05352500081062317, -0.02751999907195568, 0.0478670001...
[ -0.013879000209271908, -0.0012860000133514404, -0.005164999980479479, 0.028596999123692513, 0.016598999500274658, 0.047880999743938446, -0.06003500148653984, 0.132532998919487, 0.05372000113129616, -0.03958199918270111, -0.10492199659347534, 0.0737529993057251, 0.006643000058829784, 0.0728...
Enterprise RAG & Vector Search 2026
git clone https://github.com/langchain-ai/langgraph
Large language models (LLMs) have saturated standard medical benchmarks that test factual recall, yet their ability to perform higher-order reasoning, such as synthesizing evidence from multiple sources, remains critically under-explored.
To address this gap, we introduce MedMeta, the first benchmark designed to evaluate an LLM's ability to generate conclusions from medical meta-analyses using only the abstracts of cited studies.
MedMeta provides a challenging new benchmark for evidence synthesis and demonstrates that for clinical applications, developing robust RAG systems is a more promising direction than model specialization alone.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:50.875187
2606.31876v2
Harnessing Textual Refusal Directions for Multimodal Safety
AI Security, Red Teaming & Defense 2026
2026-06-30
[ "cs.AI", "cs.CV", "cs.LG" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
To improve safety in Large Language Models (LLMs) we can either perform post-training alignment or exploit refusal directions in the activation space. Both strategies are less feasible in Multimodal LLMs (MLLMs) as they require unsafe multimodal data
Moreno D'Incà
3
[ "Moreno D'Incà", "Nicu Sebe", "Massimiliano Mancini" ]
[]
http://arxiv.org/abs/2606.31876v2
VERIFIED_LIVE
Not Applicable
[ "https://github.com/tatsu-lab/stanford_alpaca" ]
30,247
3,991
2024-07-17
Apache-2.0
159.82
To improve safety in Large Language Models (LLMs) we can either perform post-training alignment or exploit refusal directions in the activation space. Both strategies are less feasible in Multimodal LLMs (MLLMs) as they require unsafe multimodal data, harder to collect than their unimodal counterpart. In this work, we ...
[ 0.07089299708604813, 0.046181999146938324, 0.03218900039792061, -0.05959099903702736, 0.0942780002951622, 0.0010349999647587538, -0.009296000003814697, 0.016249999403953552, 0.04644500091671944, -0.02763199992477894, 0.02811500057578087, -0.028805000707507133, 0.09411100298166275, 0.026396...
[ 0.04138699918985367, -0.12133800238370895, 0.016997000202536583, -0.04086799919605255, 0.13556599617004395, 0.029884999617934227, -0.037404000759124756, -0.013237999752163887, 0.10264299809932709, -0.042559001594781876, 0.04519300162792206, -0.05948299914598465, 0.05871700122952461, 0.0166...
AI Security, Red Teaming & Defense 2026
No code repository attached.
To improve safety in Large Language Models (LLMs) we can either perform post-training alignment or exploit refusal directions in the activation space.
Building on this, we introduce Modality-Agnostic Refusal Steering (MARS), a light-weight training-free approach that injects multimodal safety without the need for multimodal safety data.
These results reveal that safety-relevant structure is shared across modalities and that textual refusal directions are a powerful and underexplored foundation for multimodal alignment.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:22:39.184472
2606.29914v1
MemDelta: Controlled Baselines and Hidden Confounds in Agent Memory Evaluation
Enterprise RAG & Vector Search 2026
2026-06-29
[ "cs.CL", "cs.LG" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Agent memory systems are increasingly evaluated against RAG and full-context baselines, but reported gains often mix changes in the memory method with changes in the language model, embedding model, or retrieval pipeline, making it unclear what is ac
Kuan Wang
1
[ "Kuan Wang" ]
[]
http://arxiv.org/abs/2606.29914v1
VERIFIED_LIVE
https://github.com/getzep/graphiti
[ "https://github.com/getzep/graphiti" ]
29,885
3,025
2026-08-13
Apache-2.0
159.64
Agent memory systems are increasingly evaluated against RAG and full-context baselines, but reported gains often mix changes in the memory method with changes in the language model, embedding model, or retrieval pipeline, making it unclear what is actually being measured. We present MemDelta, a controlled evaluation pr...
[ 0.05411500111222267, 0.016762999817728996, -0.06276600062847137, 0.00724499998614192, -0.006955000106245279, 0.027073999866843224, -0.013760999776422977, 0.051368001848459244, 0.006370000075548887, 0.030141999945044518, -0.012137999758124352, -0.02603200078010559, 0.04363600164651871, 0.03...
[ 0.04022999852895737, -0.003379999892786145, -0.04023300111293793, 0.031971998512744904, -0.021719999611377716, 0.06476700305938721, 0.01183400023728609, 0.07612299919128418, 0.009339000098407269, -0.026952000334858894, -0.04189499840140343, -0.017153000459074974, 0.061684999614953995, 0.02...
Enterprise RAG & Vector Search 2026
git clone https://github.com/getzep/graphiti
Agent memory systems are increasingly evaluated against RAG and full-context baselines, but reported gains often mix changes in the memory method with changes in the language model, embedding model, or retrieval pipeline, making it unclear what is actually being measured.
We present MemDelta, a controlled evaluation protocol that varies one component at a time on LongMemEval-S (500 questions, 50+ sessions, three model families).
Agent memory systems are increasingly evaluated against RAG and full-context baselines, but reported gains often mix changes in the memory method with changes in the language model, embedding model, or retrieval pipeline, making it unclear what is actually being measured.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:03.072751
2605.30614v1
Audio Pirates: Black-box Audio Watermark Removal via Diffusion Priors
AI Security, Red Teaming & Defense 2026
2026-05-28
[ "cs.CR", "cs.SD" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
With the rise of AI-generated audio, watermarking has become widely used for detecting misuse and protecting intellectual property. However, adversaries may try to remove these watermarks, making it critical to evaluate how well watermarking schemes
Lingfeng Yao
9
[ "Lingfeng Yao", "Xincong Zhong", "Chenpei Huang", "Xuandong Zhao", "Hanqing Guo", "Aohan Li", "Jiang Liu", "Tomoaki Ohtsuki", "Miao Pan" ]
[]
http://arxiv.org/abs/2605.30614v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/huggingface/diffusers", "https://github.com/resemble-ai/Perth" ]
34,303
0
Unknown
Unspecified
159.53
With the rise of AI-generated audio, watermarking has become widely used for detecting misuse and protecting intellectual property. However, adversaries may try to remove these watermarks, making it critical to evaluate how well watermarking schemes withstand removal attacks. Existing attacks are often impractical: the...
[ -0.04840900003910065, 0.04705699905753136, 0.023496000096201897, -0.0709180012345314, 0.09516099840402603, -0.04396099969744682, 0.0012740000383928418, -0.11695799976587296, 0.028913000598549843, -0.06558100134134293, -0.019204000011086464, 0.06944599747657776, 0.0006549999816343188, -0.02...
[ -0.05869799852371216, -0.01215600036084652, 0.050282999873161316, -0.03697200119495392, 0.10974399745464325, -0.012463999912142754, -0.006213000044226646, -0.1212100014090538, -0.005396000109612942, -0.10305900126695633, -0.06491299718618393, -0.016126999631524086, 0.029407000169157982, -0...
AI Security, Red Teaming & Defense 2026
No code repository attached.
With the rise of AI-generated audio, watermarking has become widely used for detecting misuse and protecting intellectual property.
We propose DiffErase, a black-box watermark removal attack that assumes no knowledge of the target watermarking scheme while maintaining perceptual quality.
Theoretical analysis and extensive experiments demonstrate that inaudible audio watermarks are highly vulnerable: across multiple audio domains, DiffErase consistently removes watermarks while preserving perceptual quality.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:31:38.674161
2605.00706v1
FinSafetyBench: Evaluating LLM Safety in Real-World Financial Scenarios
AI Security, Red Teaming & Defense 2026
2026-05-01
[ "cs.CL" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Large language models (LLMs) are increasingly applied in financial scenarios. However, they may produce harmful outputs, including facilitating illegal activities or unethical behavior, posing serious compliance risks. To systematically evaluate LLM
Yutao Hou
8
[ "Yutao Hou", "Yihan Jiang", "Yuhan Xie", "Jian Yang", "Liwen Zhang", "Hailiang Huang", "Guanhua Chen", "Yun Chen" ]
[]
http://arxiv.org/abs/2605.00706v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/sustech-nlp/FinSafetyBench", "https://github.com/datalab-to/marker", "https://github.com/meta-llama/llama3" ]
38,712
0
Unknown
Unspecified
159.5
Large language models (LLMs) are increasingly applied in financial scenarios. However, they may produce harmful outputs, including facilitating illegal activities or unethical behavior, posing serious compliance risks. To systematically evaluate LLM safety in finance, we propose FinSafetyBench, a bilingual (English-Chi...
[ -0.0006200000061653554, -0.01820800080895424, -0.08448199927806854, -0.047540999948978424, 0.04784100130200386, -0.02522999979555607, 0.009712999686598778, 0.06866399943828583, 0.01775600016117096, 0.030768999829888344, -0.009592000395059586, 0.0006249999860301614, -0.019892999902367592, -...
[ -0.034046001732349396, 0.003938000183552504, -0.03818000108003616, 0.0008229999803006649, -0.009185000322759151, 0.05223799869418144, 0.010723999701440334, -0.02915400080382824, 0.08132100105285645, -0.047433000057935715, 0.01559200044721365, -0.05448099970817566, 0.014409000054001808, -0....
AI Security, Red Teaming & Defense 2026
No code repository attached.
Large language models (LLMs) are increasingly applied in financial scenarios.
To systematically evaluate LLM safety in finance, we propose FinSafetyBench, a bilingual (English-Chinese) red-teaming benchmark designed to test an LLM's refusal of requests that violate financial compliance.
Through extensive experiments on general-purpose and finance-specialized LLMs under three representative attack settings, we identify critical vulnerabilities that allow adversarial prompts to bypass compliance safeguards.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:43:21.225149
2606.25338v1
Hybrid-IR: Dual-Path Hybrid Retrieval with Iterative Reasoning for Complex Medical Question Answering
Enterprise RAG & Vector Search 2026
2026-06-24
[ "cs.CL" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Large language models (LLMs) have shown promising performance across a wide range of biomedical applications, including medical question answering (QA), yet they remain prone to hallucinations and outdated knowledge. Although retrieval-augmented gene
Sheng Wan
4
[ "Sheng Wan", "Jiahui Zhang", "Zicheng Zhao", "Shougang Ren" ]
[]
http://arxiv.org/abs/2606.25338v1
VERIFIED_LIVE
https://github.com/meta-llama/llama3
[ "https://github.com/meta-llama/llama3" ]
29,261
0
Unknown
Unspecified
159.16
Large language models (LLMs) have shown promising performance across a wide range of biomedical applications, including medical question answering (QA), yet they remain prone to hallucinations and outdated knowledge. Although retrieval-augmented generation (RAG) can alleviate this issue by incorporating external docume...
[ -0.009643999859690666, 0.04203899949789047, 0.02898300066590309, 0.02550400048494339, -0.04774300009012222, -0.03153200075030327, 0.07902900129556656, 0.12638500332832336, 0.07044900208711624, -0.05543600022792816, -0.06983000040054321, 0.017726000398397446, -0.012107999995350838, 0.118277...
[ -0.020555999130010605, 0.03606000170111656, 0.005135999992489815, 0.015660999342799187, -0.049981001764535904, 0.021591000258922577, -0.0067819999530911446, 0.07185199856758118, 0.02918200008571148, -0.06088100001215935, -0.050216998904943466, 0.03207800164818764, 0.022809000685811043, 0.0...
Enterprise RAG & Vector Search 2026
git clone https://github.com/meta-llama/llama3
Large language models (LLMs) have shown promising performance across a wide range of biomedical applications, including medical question answering (QA), yet they remain prone to hallucinations and outdated knowledge.
First, medical knowledge is often fragmented across documents, while most RAG methods rely on a single retrieval path, which makes it challenging to jointly preserve fine-grained semantic information and structured global associations.
Experiments on three widely used medical QA benchmarks demonstrate the effectiveness of our Hybrid-IR.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:07.499451
2605.05509v1
WAAA! Web Adversaries Against Agentic Browsers
AI Security, Red Teaming & Defense 2026
2026-05-06
[ "cs.CR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Large language models (LLMs) are increasingly being integrated into web browsers to create agentic browsing systems that execute actions on behalf of the user. Prior work considering the security of agentic browsers focuses exclusively on indirect pr
Sohom Datta
4
[ "Sohom Datta", "Alex Nahapetyan", "William Enck", "Alexandros Kapravelos" ]
[]
http://arxiv.org/abs/2605.05509v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/features/copilot", "https://github.com/microsoft/playwright-mcp" ]
36,079
0
Unknown
Unspecified
158.98
Large language models (LLMs) are increasingly being integrated into web browsers to create agentic browsing systems that execute actions on behalf of the user. Prior work considering the security of agentic browsers focuses exclusively on indirect prompt-injection attacks. However, by failing to consider traditional we...
[ -0.09682799875736237, 0.016733000054955482, -0.04606400057673454, -0.07054600119590759, -0.0026229999493807554, -0.07003100216388702, 0.03581799939274788, -0.07193200290203094, -0.0248429998755455, -0.002176000038161874, 0.028812000527977943, -0.0002880000101868063, 0.04331599920988083, 0....
[ -0.05847200006246567, -0.04594599828124046, -0.020159000530838966, -0.011274999938905239, 0.026686999946832657, -0.010254000313580036, 0.0440870001912117, -0.05374300107359886, 0.019929999485611916, -0.04138600081205368, 0.0018220000201836228, -0.023264000192284584, 0.11389599740505219, 0....
AI Security, Red Teaming & Defense 2026
No code repository attached.
Large language models (LLMs) are increasingly being integrated into web browsers to create agentic browsing systems that execute actions on behalf of the user.
However, by failing to consider traditional web attacks, previous agentic browser threat models have a blind spot to web social engineering attacks originally designed to trick humans.
We show that agentic browsers exhibit five major failure modes when facing traditional and LLM web threats, demonstrating the need to rearchitect agentic browsers before they are ready for the current web.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:41:32.348563
2605.14309v2
ICED: Concept-level Machine Unlearning via Interpretable Concept Decomposition
AI Security, Red Teaming & Defense 2026
2026-05-14
[ "cs.CV", "cs.AI", "cs.LG" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Machine unlearning in Vision-Language Models (VLMs) is typically performed at the image or instance level, making it difficult to precisely remove target knowledge without affecting unrelated semantics. This issue is especially pronounced since a sin
Shen Lin
5
[ "Shen Lin", "Jing Lin", "Junhao Dong", "Piotr Koniusz", "Li Xu" ]
[]
http://arxiv.org/abs/2605.14309v2
VERIFIED_LIVE
Not Applicable
[ "https://github.com/openai/CLIP" ]
34,167
0
Unknown
Unspecified
158.79
Machine unlearning in Vision-Language Models (VLMs) is typically performed at the image or instance level, making it difficult to precisely remove target knowledge without affecting unrelated semantics. This issue is especially pronounced since a single image often contains multiple entangled concepts, including both t...
[ -0.0030239999759942293, -0.052703000605106354, -0.013458999805152416, 0.06273499876260757, 0.06768299639225006, -0.09211000055074692, 0.01309099979698658, -0.051614001393318176, -0.06186800077557564, 0.0047439998015761375, -0.07006499916315079, -0.1131879985332489, 0.09763800352811813, 0.0...
[ 0.07830999791622162, -0.08501199632883072, -0.04197600111365318, 0.0822179988026619, 0.10191299766302109, 0.0122060002759099, 0.010641000233590603, -0.06619399785995483, -0.01002500019967556, -0.04674199968576431, -0.013113000430166721, -0.0479310005903244, 0.07528200000524521, 0.065682001...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Machine unlearning in Vision-Language Models (VLMs) is typically performed at the image or instance level, making it difficult to precisely remove target knowledge without affecting unrelated semantics.
In this paper, we propose an interpretable concept-level unlearning framework for VLMs, which constructs a compact task-specific concept vocabulary from the forgetting set using a multimodal large language model.
Extensive experiments across both in-domain and out-of-domain forgetting settings demonstrate that our method enables more comprehensive target forgetting, better preserves non-target knowledge within the same image, and maintains competitive model utility compared with existing VLM unlearning methods.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:37:36.842296
2606.07970v1
Defending Against Malicious Finetuning by Scaling Train-time Adversarial Attacks
AI Security, Red Teaming & Defense 2026
2026-06-06
[ "cs.CL", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Current open-weight large language models (LLMs) are prone to malicious finetuning attacks, which could compromise the safety alignment of LLMs with only a few steps of supervised finetuning (SFT) on poisoned datasets. Existing alignment-stage defens
Haoming Wen
7
[ "Haoming Wen", "Shi Chen", "Qingyu Shi", "Siyuan Liu", "Minrui Luo", "Jingzhao Zhang", "Tianxing He" ]
[]
http://arxiv.org/abs/2606.07970v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/haomingwen/patcher", "https://github.com/tatsu-lab/stanford_alpaca" ]
30,247
3,991
2024-07-17
Apache-2.0
158.62
Current open-weight large language models (LLMs) are prone to malicious finetuning attacks, which could compromise the safety alignment of LLMs with only a few steps of supervised finetuning (SFT) on poisoned datasets. Existing alignment-stage defenses are primarily designed to defend against attacks that use parameter...
[ -0.09310100227594376, -0.03017199970781803, 0.04632800072431564, 0.04876700043678284, 0.092958003282547, 0.02768000029027462, 0.040045998990535736, -0.025158999487757683, -0.06524299830198288, -0.05553799867630005, 0.02004300057888031, -0.021625999361276627, -0.012409999966621399, -0.04789...
[ -0.05364200100302696, -0.0603410005569458, 0.019588999450206757, 0.02552499994635582, 0.05889499932527542, 0.02157999947667122, -0.01826700009405613, -0.02032499946653843, -0.016349999234080315, -0.029678000137209892, 0.01526500005275011, -0.07346799969673157, -0.007747000083327293, -0.017...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Current open-weight large language models (LLMs) are prone to malicious finetuning attacks, which could compromise the safety alignment of LLMs with only a few steps of supervised finetuning (SFT) on poisoned datasets.
Existing alignment-stage defenses are primarily designed to defend against attacks that use parameter-efficient finetuning methods.
Extensive experiments show that Patcher substantially improves the model's robustness compared to vanilla SFT alignment, and transfers to diverse attack scenarios and model sizes.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:28:22.573930
2606.07963v1
Shared Latent Structures Enable Unified Backdoor Detection and Mitigation in LLMs
AI Security, Red Teaming & Defense 2026
2026-06-06
[ "cs.AI", "cs.CL" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Backdoor attacks in large language models (LLMs) are often treated as isolated trigger-response failures, motivating defenses tailored to specific triggers or behaviors. We show this view is incomplete. Across diverse backdoor behaviors, we identify
Omar Mahmoud
7
[ "Omar Mahmoud", "Aly M. Kassem", "Thommen George Karimpanal", "Buddhika Laknath Semage", "Negar Rostamzadeh", "Golnoosh Farnadi", "Santu Rana" ]
[]
http://arxiv.org/abs/2606.07963v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/tatsu-lab/stanford_alpaca" ]
30,247
3,991
2024-07-17
Apache-2.0
158.62
Backdoor attacks in large language models (LLMs) are often treated as isolated trigger-response failures, motivating defenses tailored to specific triggers or behaviors. We show this view is incomplete. Across diverse backdoor behaviors, we identify a shared latent mechanism that can be detected, causally controlled, a...
[ -0.012509999796748161, -0.022390000522136688, 0.0016870000399649143, -0.012242999859154224, 0.1222900003194809, -0.023615000769495964, -0.05747000128030777, -0.08175099641084671, 0.013237000443041325, 0.026820000261068344, 0.05544700101017952, 0.022947000339627266, 0.06550200283527374, 0.0...
[ -0.03853800147771835, -0.0832739993929863, 0.025899000465869904, 0.04496899992227554, 0.08865199983119965, 0.057231999933719635, -0.0019519999623298645, -0.10025300085544586, 0.06303700059652328, -0.00922400038689375, -0.013282000087201595, -0.05069199949502945, 0.06373000144958496, -0.015...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Backdoor attacks in large language models (LLMs) are often treated as isolated trigger-response failures, motivating defenses tailored to specific triggers or behaviors.
Across diverse backdoor behaviors, we identify a shared latent mechanism that can be detected, causally controlled, and suppressed.
Together, our results suggest that many backdoors rely on a transferable latent mechanism, enabling unified detection and mitigation.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:28:23.722178
2606.05817v1
Consistency Training Along the Transformer Stack
AI Security, Red Teaming & Defense 2026
2026-06-04
[ "cs.LG", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Consistency training encourages models to behave similarly across different contexts, and has shown promise for reducing misalignment. We broaden the scope of consistency training in two ways. First, we introduce two new internal consistency targets:
Sukrati Gautam
10
[ "Sukrati Gautam", "Neil Shah", "Arav Dhoot", "Bryan Maruyama", "Caroline Wei", "Rohan Kapoor", "Robert Sidey", "Prakhar Gupta", "Zi Cheng Huang", "David Demitri Africa" ]
[]
http://arxiv.org/abs/2606.05817v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/tatsu-lab/stanford_alpaca", "https://github.com/c-wei/AttCT" ]
30,247
3,991
2024-07-17
Apache-2.0
158.52
Consistency training encourages models to behave similarly across different contexts, and has shown promise for reducing misalignment. We broaden the scope of consistency training in two ways. First, we introduce two new internal consistency targets: MLP Consistency Training (MLPCT), which matches post-activation MLP s...
[ -0.14002999663352966, -0.08086399734020233, 0.023538000881671906, 0.005793000105768442, -0.02662999927997589, -0.03152500092983246, -0.060451000928878784, 0.0017330000409856439, -0.06341300159692764, -0.11949200183153152, -0.03170600160956383, 0.04556399956345558, 0.044530998915433884, 0.0...
[ -0.0285630002617836, -0.11065000295639038, 0.011043000034987926, 0.0016919999616220593, 0.05147600173950195, 0.051892999559640884, 0.034338999539613724, -0.053741998970508575, 0.015173999592661858, -0.10479699820280075, 0.03608199954032898, -0.06587900221347809, 0.05337899923324585, -0.002...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Consistency training encourages models to behave similarly across different contexts, and has shown promise for reducing misalignment.
First, we introduce two new internal consistency targets: MLP Consistency Training (MLPCT), which matches post-activation MLP states, and Attention Consistency Training (AttCT), which matches per-head attention distributions.
Our results suggest that consistency training is a flexible and extensible framework for alignment, capable of unifying defenses against a broader class of model pathologies.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:28:54.902234
2606.04168v1
When Autoregressive Consistency Hurts Safety Alignment
AI Security, Red Teaming & Defense 2026
2026-06-02
[ "cs.LG", "cs.CR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Safety alignment in large language models (LLMs) is fragile in part because it is often shallow: fine-tuning mainly reshapes the model's behavior near the first few output tokens. We argue that this phenomenon can be understood through autoregressive
Bochen Lyu
4
[ "Bochen Lyu", "Yiyang Jia", "Xiaohao Cai", "Zhanxing Zhu" ]
[]
http://arxiv.org/abs/2606.04168v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/tatsu-lab/stanford_alpaca" ]
30,247
3,991
2024-07-17
Apache-2.0
158.42
Safety alignment in large language models (LLMs) is fragile in part because it is often shallow: fine-tuning mainly reshapes the model's behavior near the first few output tokens. We argue that this phenomenon can be understood through autoregressive consistency, the tendency of next-token prediction to preserve and ex...
[ -0.14038999378681183, -0.07090800255537033, -0.0033829999156296253, 0.021486999467015266, 0.007848000153899193, 0.05086199939250946, -0.005847000051289797, -0.011039000004529953, 0.04972099885344505, -0.018675999715924263, 0.02415899932384491, 0.0009640000062063336, 0.0049680001102387905, ...
[ -0.07821500301361084, -0.05934999883174896, 0.004972000140696764, -0.008747999556362629, 0.06655500084161758, 0.04826899990439415, -0.017481999471783638, -0.026693999767303467, 0.09629400074481964, 0.0070130000822246075, -0.006254000123590231, -0.022321999073028564, 0.050714001059532166, -...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Safety alignment in large language models (LLMs) is fragile in part because it is often shallow: fine-tuning mainly reshapes the model's behavior near the first few output tokens.
The same mechanism also predicts a broader class of attacks on LLMs: attacks that induce harmful continuation states at arbitrary positions in the output trajectory.
Overall, our results suggest that autoregressive consistency should be treated as a central consideration in both safety alignment and attack design.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:29:37.543792
2606.09441v1
SIFT: Selective-Index For Fast Compute of RAG Prefill by Exploiting Attention Invariance
Enterprise RAG & Vector Search 2026
2026-06-08
[ "cs.AI", "cs.AR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Retrieval-Augmented Generation (RAG) injects LLM queries with relevant documents to improve response quality. This injection increases prompt length and slows time to first token (TTFT). Unlike standard queries, RAG queries have a unique property of
Rya Sanovar
4
[ "Rya Sanovar", "Srikant Bharadwaj", "Hritvik Taneja", "Moinuddin Qureshi" ]
[]
http://arxiv.org/abs/2606.09441v1
VERIFIED_LIVE
https://github.com/meta-llama/llama3
[ "https://github.com/meta-llama/llama3" ]
29,261
0
Unknown
Unspecified
158.36
Retrieval-Augmented Generation (RAG) injects LLM queries with relevant documents to improve response quality. This injection increases prompt length and slows time to first token (TTFT). Unlike standard queries, RAG queries have a unique property of context reuse where the same documents recur across user queries. Thus...
[ -0.050515998154878616, 0.04677800089120865, 0.05449000000953674, -0.007472999859601259, -0.05025799944996834, 0.011175000108778477, 0.08749199658632278, -0.014851000159978867, 0.005483000073581934, -0.039000000804662704, -0.02227799966931343, 0.011068999767303467, -0.029069000855088234, -0...
[ -0.021956000477075577, -0.025875000283122063, 0.03151499852538109, 0.08825899660587311, 0.03786800056695938, -0.020752999931573868, 0.018533999100327492, -0.01568399928510189, 0.031686000525951385, -0.05303199961781502, -0.01235199999064207, 0.042555999010801315, 0.01450899988412857, -0.06...
Enterprise RAG & Vector Search 2026
git clone https://github.com/meta-llama/llama3
Retrieval-Augmented Generation (RAG) injects LLM queries with relevant documents to improve response quality.
This injection increases prompt length and slows time to first token (TTFT).
During prefill, SIFT computes the attention only for the marked locations and improves TTFT by 1.71x while holding accuracy within 1% of full recompute.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:21.081648
2606.04231v1
MM-BizRAG: Rethinking Multimodal Retrieval-Augmented Generation for General Purpose Enterprise Q&A
Enterprise RAG & Vector Search 2026
2026-06-02
[ "cs.CL", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Recent advances in multimodal retrieval-augmented generation (MM-RAG) have shifted toward minimal parsing, relying on page-level images for producing retriever embeddings and for answer generation. While efficient, this trend often neglects explicit
Hanoz Bhathena
10
[ "Hanoz Bhathena", "Parin Rajesh Jhaveri", "Rohan Mittal", "Prateek Singh", "Aymen Kallala", "Rachneet Kaur", "Yiqiao Jin", "Zhen Zeng", "Adwait Ratnaparkhi", "Denis Kochedykov" ]
[]
http://arxiv.org/abs/2606.04231v1
VERIFIED_LIVE
https://github.com/JaidedAI/EasyOCR
[ "https://github.com/JaidedAI/EasyOCR", "https://github.com/pypdfium2-team/pypdfium2", "https://github.com/vibrantlabsai/ragas" ]
29,905
3,597
2025-12-05
Apache-2.0
158.29
Recent advances in multimodal retrieval-augmented generation (MM-RAG) have shifted toward minimal parsing, relying on page-level images for producing retriever embeddings and for answer generation. While efficient, this trend often neglects explicit handling of the rich, structured information in complex enterprise doc...
[ -0.03360399976372719, -0.04709399864077568, 0.005642999894917011, 0.0004189999890513718, 0.024396000429987907, 0.05012200027704239, 0.007331999950110912, -0.018703000620007515, 0.06750600039958954, -0.058045998215675354, 0.008236000314354897, -0.012753999792039394, 0.04227200150489807, 0.0...
[ -0.055417001247406006, -0.003352999920025468, 0.051683999598026276, 0.03243599832057953, 0.03744100034236908, 0.016464000567793846, -0.07591000199317932, 0.01773899979889393, 0.05501899868249893, -0.0641229972243309, -0.04830300062894821, 0.0370590016245842, 0.06105300039052963, 0.04867500...
Enterprise RAG & Vector Search 2026
git clone https://github.com/JaidedAI/EasyOCR
Recent advances in multimodal retrieval-augmented generation (MM-RAG) have shifted toward minimal parsing, relying on page-level images for producing retriever embeddings and for answer generation.
In this work, we take a more direct approach: MM-BizRAG proactively extracts and represents document structure via a document structure-aware split that dynamically routes documents through orientation-specific ingestion pipelines, applying explicit layout-aware parsing for vertically structured documents (e.g., report...
Through experiments on a large, heterogeneous enterprise dataset and two public benchmarks (SlideVQA and FinRAGBench-V), MM-BizRAG consistently outperforms state-of-the-art vision-centric baselines by up to 32% points, with especially strong gains on report-style layouts.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:29.868336
2605.28467v1
Mitigating Adaptive Attacks against Reasoning Models with Activation Consistency Training
AI Security, Red Teaming & Defense 2026
2026-05-27
[ "cs.LG" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
As LLMs gain stronger reasoning capabilities, their extended chain-of-thought introduces new degrees of complexity for defending against adversarial jailbreaks and prompt injection. We study consistency training, a family of fine-tuning objectives th
Avidan Shah
3
[ "Avidan Shah", "Jannik Brinkmann", "Rico Angell" ]
[]
http://arxiv.org/abs/2605.28467v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/tatsu-lab/stanford_alpaca" ]
30,247
3,991
2024-07-17
Apache-2.0
158.12
As LLMs gain stronger reasoning capabilities, their extended chain-of-thought introduces new degrees of complexity for defending against adversarial jailbreaks and prompt injection. We study consistency training, a family of fine-tuning objectives that enforce identical behavior on clean prompts and adversarial rewrite...
[ -0.04653799906373024, -0.07534600049257278, 0.019917000085115433, 0.1022069975733757, 0.016385000199079514, 0.014183999970555305, 0.043366000056266785, -0.027720000594854355, 0.008023000322282314, -0.03851599991321564, 0.003389999968931079, -0.054993998259305954, 0.11100800335407257, 0.026...
[ -0.016607999801635742, -0.09883499890565872, 0.005016000010073185, 0.016534000635147095, 0.04848900064826012, 0.02955400012433529, 0.04326599836349487, -0.028667999431490898, -0.018403999507427216, 0.02174600027501583, -0.0182460006326437, -0.013055999763309956, 0.0531340017914772, 0.00056...
AI Security, Red Teaming & Defense 2026
No code repository attached.
As LLMs gain stronger reasoning capabilities, their extended chain-of-thought introduces new degrees of complexity for defending against adversarial jailbreaks and prompt injection.
We formulate both methods as a prompt injection defense and find ACT to be competitive with other training-based defenses while requiring only self-supervised pairs of clean and wrapped prompts.
Together, these results suggest that supervising internal representations is a surprisingly effective and interpretable approach to various forms of safety training in reasoning models.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:32:32.257285
2606.04442v1
MemoryDocDataSet: A Benchmark for Joint Conversational Memory and Long Document Reasoning
Enterprise RAG & Vector Search 2026
2026-06-03
[ "cs.CL", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
AI systems increasingly need to combine two demanding capabilities: navigating multi-session conversation history and performing deep reading comprehension within long documents. Yet no existing benchmark evaluates both simultaneously. We introduce M
Qiyang Xie
7
[ "Qiyang Xie", "Jialun Wu", "Xinjie He", "Su Liu", "Shuai Xiao", "Zhiyuan Lin", "Weikai Zhou" ]
[]
http://arxiv.org/abs/2606.04442v1
VERIFIED_LIVE
https://github.com/chroma-core/chroma
[ "https://github.com/chroma-core/chroma" ]
29,050
0
Unknown
Unspecified
158.03
AI systems increasingly need to combine two demanding capabilities: navigating multi-session conversation history and performing deep reading comprehension within long documents. Yet no existing benchmark evaluates both simultaneously. We introduce MemoryDocDataSet, a synthetic benchmark of 50 micro-worlds and 1,000 QA...
[ 0.033553000539541245, -0.025736000388860703, -0.027594000101089478, 0.03657799959182739, 0.05482799932360649, -0.028111999854445457, 0.017805000767111778, 0.06132499873638153, 0.053619999438524246, -0.006442999932914972, 0.004447000101208687, 0.019756000488996506, -0.028024999424815178, 0....
[ 0.01992199942469597, -0.02265400066971779, -0.021953999996185303, 0.028384000062942505, 0.04224200174212456, -0.0045529999770224094, -0.02261899970471859, 0.05061500146985054, 0.01883699931204319, -0.017578000202775, -0.06547000259160995, -0.033934999257326126, -0.012047999538481236, 0.031...
Enterprise RAG & Vector Search 2026
git clone https://github.com/chroma-core/chroma
AI systems increasingly need to combine two demanding capabilities: navigating multi-session conversation history and performing deep reading comprehension within long documents.
We introduce MemoryDocDataSet, a synthetic benchmark of 50 micro-worlds and 1,000 QA pairs in which each instance comprises 3-5 personas, a temporal event graph spanning months of activity, 3-5 real long documents (20,000-50,000 tokens each sourced from the Caselaw Access Project), multi-session conversations grounded ...
The best baseline (RAG-Both) achieves 0.358 overall F1 and 0.342 on Hybrid.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:26.837241
2605.28116v1
MIRAGE: Context-Aware Prompt Injection against Mobile GUI Agents via User-Generated Content
AI Security, Red Teaming & Defense 2026
2026-05-27
[ "cs.CR", "cs.AI", "cs.CL" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Mobile graphical user interface (GUI) agents driven by vision-language models (VLMs) perceive the screen as rendered pixels and choose actions from what they see, so they cannot reliably separate trusted interface elements from user-generated content
Ruoqi Guo
10
[ "Ruoqi Guo", "Yi Liu", "Gelei Deng", "Yiheng Xiong", "Yuekang Li", "Ying Zhang", "Leo Yu Zhang", "Lida Zhao", "Ji Jie", "Yuxiao Lu" ]
[]
http://arxiv.org/abs/2605.28116v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/JaidedAI/EasyOCR" ]
29,905
0
Unknown
Unspecified
157.99
Mobile graphical user interface (GUI) agents driven by vision-language models (VLMs) perceive the screen as rendered pixels and choose actions from what they see, so they cannot reliably separate trusted interface elements from user-generated content. We present MIRAGE (Mobile Injection of Realistic Adversarial GUI Exa...
[ -0.007404000032693148, 0.05806000158190727, 0.005235000047832727, -0.05593400076031685, 0.0509909987449646, -0.057468000799417496, 0.0950310006737709, -0.007433000020682812, 0.015031999908387661, -0.035273000597953796, 0.02325800061225891, -0.09044499695301056, 0.134785994887352, 0.0065469...
[ -0.02673500031232834, 0.01882000081241131, -0.00267699989490211, -0.04441399872303009, 0.10307999700307846, -0.04449699819087982, 0.03893600031733513, -0.050923001021146774, 0.016102999448776245, -0.05049800127744675, 0.010871999897062778, -0.11536300182342529, 0.12510600686073303, 0.00976...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Mobile graphical user interface (GUI) agents driven by vision-language models (VLMs) perceive the screen as rendered pixels and choose actions from what they see, so they cannot reliably separate trusted interface elements from user-generated content.
We present MIRAGE (Mobile Injection of Realistic Adversarial GUI Examples), a pipeline that turns benign mobile screenshots into prompt-injection samples by placing attacker-controlled text into ordinary user-generated content regions, without modifying the agent, the application, or the operating system.
On a 1,111-sample benchmark spanning ten applications and eleven attack intents, all five evaluated VLM agents are vulnerable, with attack success rates of 23%-30%, and MIRAGE scores higher on human realism ratings than the strongest prior attack (3.02 versus 2.52 out of 5).
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:32:38.447390
2605.21706v2
Latent-space Attacks for Refusal Evasion in Language Models
AI Security, Red Teaming & Defense 2026
2026-05-20
[ "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Safety-aligned language models are trained to refuse harmful requests, yet refusal behavior can be suppressed by steering their internal representations. Existing methods do so by ablating a refusal direction from model activations, aiming to remove
Giorgio Piras
7
[ "Giorgio Piras", "Raffaele Mura", "Fabio Brau", "Maura Pintor", "Luca Oneto", "Fabio Roli", "Battista Biggio" ]
[]
http://arxiv.org/abs/2605.21706v2
VERIFIED_LIVE
Not Applicable
[ "https://github.com/tatsu-lab/stanford_alpaca", "https://github.com/pralab/latent-evasion" ]
30,247
3,991
2024-07-17
Apache-2.0
157.77
Safety-aligned language models are trained to refuse harmful requests, yet refusal behavior can be suppressed by steering their internal representations. Existing methods do so by ablating a refusal direction from model activations, aiming to remove refusal from the model's residual stream. Despite their empirical succ...
[ -0.023060999810695648, 0.016332000494003296, 0.06947100162506104, 0.016172999516129494, 0.03962999954819679, 0.01114100031554699, 0.08423899859189987, -0.08964300155639648, 0.14407700300216675, -0.019951000809669495, 0.08419699966907501, -0.055121999233961105, 0.11496700346469879, -0.01366...
[ -0.02438800036907196, -0.011010999791324139, 0.05018499866127968, 0.02872299961745739, 0.06137700006365776, -0.009840000420808792, 0.031637001782655716, -0.07726199924945831, 0.12870100140571594, -0.013206999748945236, 0.00031999999191612005, -0.05930500105023384, 0.10637900233268738, -0.0...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Safety-aligned language models are trained to refuse harmful requests, yet refusal behavior can be suppressed by steering their internal representations.
Existing methods do so by ablating a refusal direction from model activations, aiming to remove refusal from the model's residual stream.
We achieve state-of-the-art attack success rate across 15 instruction-tuned, multimodal, and reasoning models, outperforming existing refusal-ablation baselines and specialized jailbreak attacks.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:34:36.520083
2605.18309v1
Alignment Dynamics in LLM Fine-Tuning
AI Security, Red Teaming & Defense 2026
2026-05-18
[ "cs.LG", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Although Large Language Models (LLMs) achieve strong alignment through supervised fine-tuning and reinforcement learning from human feedback, the alignment is often fragile under subsequent fine-tuning. Existing explanations either attribute alignmen
Yuhan Huang
3
[ "Yuhan Huang", "Huanran Chen", "Yinpeng Dong" ]
[]
http://arxiv.org/abs/2605.18309v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/tatsu-lab/stanford_alpaca" ]
30,247
3,991
2024-07-17
Apache-2.0
157.67
Although Large Language Models (LLMs) achieve strong alignment through supervised fine-tuning and reinforcement learning from human feedback, the alignment is often fragile under subsequent fine-tuning. Existing explanations either attribute alignment fragility to gradient geometry or characterize it as a distributiona...
[ 0.02185799926519394, -0.10763700306415558, -0.011436999775469303, -0.06151000037789345, -0.07835700362920761, -0.031644999980926514, -0.061535999178886414, -0.016669999808073044, 0.013774000108242035, -0.014161000028252602, 0.001740999985486269, 0.030990000814199448, -0.006099000107496977, ...
[ -0.020262999460101128, -0.16224199533462524, 0.04987499862909317, 0.015250000171363354, 0.024306999519467354, 0.08806099742650986, -0.030804000794887543, -0.057472001761198044, 0.11088299751281738, -0.0018400000408291817, -0.019511999562382698, 0.011552000418305397, -0.018219999969005585, ...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Although Large Language Models (LLMs) achieve strong alignment through supervised fine-tuning and reinforcement learning from human feedback, the alignment is often fragile under subsequent fine-tuning.
In this work, we introduce a tractable alignment score and derive its closed-form update during fine-tuning, yielding a unified framework for alignment dynamics.
Together, these results provide a unified dynamical perspective on how alignment is disrupted and reactivated during LLM fine-tuning.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:35:39.714430
2606.15631v1
Retrieve, Don't Retrain: Extending Vision Language Action Models to New Tasks at Test Time
Enterprise RAG & Vector Search 2026
2026-06-14
[ "cs.RO", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Extending a vision-language-action (VLA) policy to a new task typically requires task-specific teleoperated demonstrations and per-task fine-tuning, making adaptation costly in both data collection and compute. In this paper, we show that this target
Jeongeun Park
6
[ "Jeongeun Park", "Juhan Park", "Taekyung Kim", "Sungjoon Choi", "Dongyoon Han", "Sangdoo Yun" ]
[]
http://arxiv.org/abs/2606.15631v1
VERIFIED_LIVE
https://github.com/huggingface/lerobot
[ "https://github.com/huggingface/lerobot" ]
26,626
0
Unknown
Unspecified
157.63
Extending a vision-language-action (VLA) policy to a new task typically requires task-specific teleoperated demonstrations and per-task fine-tuning, making adaptation costly in both data collection and compute. In this paper, we show that this target-side per-task adaptation cost can be replaced by retrieval. Our retri...
[ -0.025398999452590942, -0.11659400165081024, 0.0176790002733469, 0.043512001633644104, 0.0682229995727539, 0.044075001031160355, 0.052932001650333405, -0.03514299914240837, -0.025141999125480652, -0.04779199883341789, 0.022971000522375107, -0.05313900113105774, -0.011433999985456467, 0.095...
[ -0.04648499935865402, -0.08393599838018417, 0.05045900121331215, -0.03860900178551674, 0.047759998589754105, 0.05064700171351433, 0.034040000289678574, -0.03585200011730194, 0.015893999487161636, -0.009052000008523464, 0.002093000104650855, -0.06637799739837646, -0.0023089998867362738, 0.0...
Enterprise RAG & Vector Search 2026
git clone https://github.com/huggingface/lerobot
Extending a vision-language-action (VLA) policy to a new task typically requires task-specific teleoperated demonstrations and per-task fine-tuning, making adaptation costly in both data collection and compute.
In this paper, we show that this target-side per-task adaptation cost can be replaced by retrieval.
On PushT, we study how retrieval provides a reusable high-level motion prior for cross-embodiment generalization to unseen goal angles, while on RoboTwin 2.0 our method outperforms cross-embodiment baselines on unseen tasks, and we additionally demonstrate the method on a real robot.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:18.951663
2605.24958v1
SEP-Attack: A Simple and Effective Paradigm for Transfer-Based Textual Adversarial Attack
AI Security, Red Teaming & Defense 2026
2026-05-24
[ "cs.CL", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Despite the strong performance of deep neural networks in modern Web and language applications, they remain vulnerable to adversarial attacks, especially transferable attacks that generate adversarial examples using surrogate models without accessing
Han Liu
8
[ "Han Liu", "Zhi Xu", "Xiaotong Zhang", "Feng Zhang", "Xiaoming Xu", "Wei Wang", "Fenglong Ma", "Hong Yu" ]
[]
http://arxiv.org/abs/2605.24958v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/meta-llama/llama3" ]
29,261
0
Unknown
Unspecified
157.61
Despite the strong performance of deep neural networks in modern Web and language applications, they remain vulnerable to adversarial attacks, especially transferable attacks that generate adversarial examples using surrogate models without accessing the victim model. Transferable attacks in the text domain are still u...
[ -0.0756239965558052, 0.003358999965712428, -0.01575000025331974, 0.0028510000556707382, 0.10572099685668945, 0.02338000014424324, 0.03630400076508522, -0.07517000287771225, 0.05222000181674957, -0.03901199996471405, -0.02094399929046631, -0.048548001796007156, 0.06663300096988678, -0.02969...
[ -0.07015799731016159, -0.09551999717950821, 0.014383000321686268, 0.03521300107240677, 0.1318340003490448, 0.02743300050497055, 0.01680699922144413, -0.038624998182058334, 0.014670000411570072, -0.035576000809669495, -0.007832000032067299, -0.055413998663425446, 0.04198399931192398, 0.0094...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Despite the strong performance of deep neural networks in modern Web and language applications, they remain vulnerable to adversarial attacks, especially transferable attacks that generate adversarial examples using surrogate models without accessing the victim model.
To address these challenges, we propose a simple yet effective paradigm for transfer-based textual adversarial attack, named SEP-Attack.
Experiments conducted on four datasets and two real-world APIs validate the efficacy of SEP-Attack, significantly outperforming state-of-the-art baselines.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:33:44.605845
2608.09290v2
OpenCodeReview: Determinism over Non-Determinism for Cost-Effective Agent-Based Code Review
Autonomous AI Agents & Swarms 2026
2026-08-10
[ "cs.SE" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
LLM-based code review agents promise scalable, always-on review, yet current systems suffer from two intertwined weaknesses: (1) non-determinism--unbounded tool use makes review outcomes unstable, and (2) context locality--the reviewer's access remai
Zhengfeng Li
11
[ "Zhengfeng Li", "Lei Zhang", "Xianwei Wu", "Zhengqi Zhuang", "Yingjie Xu", "Boge Wang", "Shaofei Zhu", "Chuan Wang", "Peng Zhao", "Xinyu Zheng", "Guoping Rong" ]
[ "Alibaba Group" ]
http://arxiv.org/abs/2608.09290v2
VERIFIED_LIVE
https://github.com/alibaba/open-code-review
[ "https://github.com/alibaba/open-code-review" ]
20,650
0
Unknown
Unspecified
157.52
LLM-based code review agents promise scalable, always-on review, yet current systems suffer from two intertwined weaknesses: (1) non-determinism--unbounded tool use makes review outcomes unstable, and (2) context locality--the reviewer's access remains bounded to the diff, capping discoverable issue depth. Both give ri...
[ -0.02571200020611286, -0.020723000168800354, -0.1330839991569519, 0.06809200346469879, 0.038899000734090805, -0.009310999885201454, -0.0071669998578727245, -0.049157001078128815, 0.010979999788105488, 0.024511000141501427, 0.0024800000246614218, -0.0178849995136261, 0.020569000393152237, -...
[ -0.09384600073099136, -0.043609000742435455, -0.10136300325393677, 0.07402999699115753, 0.07184500247240067, -0.06302899867296219, -0.012698999606072903, -0.015407999977469444, 0.01586800068616867, 0.019740000367164612, -0.023541999980807304, -0.027303999289870262, 0.03233100101351738, -0....
Autonomous AI Agents & Swarms 2026
git clone https://github.com/alibaba/open-code-review && cd open-code-review && (pip install -e . || pip install -r requirements.txt)
LLM-based code review agents promise scalable, always-on review, yet current systems suffer from two intertwined weaknesses: (1) non-determinism--unbounded tool use makes review outcomes unstable, and (2) context locality--the reviewer's access remains bounded to the diff, capping discoverable issue depth.
To address these, we introduce OpenCodeReview, built on deterministic engineering for uncertain agents: rather than granting maximal freedom, we inject determinism at three deliberate pipeline points.
On AACR-Bench (200 real-world PRs, 10 languages, 1,505 expert-verified comments), OpenCodeReview outperforms mainstream coding agents (e.g., Claude Code and Codex) across six LLM backends, achieving up to 2.17x higher SEM-F1 (25.10% vs. 11.57%) while consuming 5-15x fewer tokens.
1
Autonomous AI Agents & Swarms 2026 Research
Explosive (>50/mo)
787
2026-08-17T15:54:50.596909
2605.22984v1
Test-Time Training Undermines Safety Guardrails
AI Security, Red Teaming & Defense 2026
2026-05-21
[ "cs.LG", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Test-Time Training (TTT) is an emerging paradigm that enables models to adapt their parameters during inference, improving performance on tasks such as few-shot learning, retrieval-augmented generation, and complex reasoning. However, this dynamic ad
Simone Antonelli
3
[ "Simone Antonelli", "Sadegh Akhondzadeh", "Aleksandar Bojchevski" ]
[]
http://arxiv.org/abs/2605.22984v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/uoc-tail/ttt-jailbreak", "https://github.com/huggingface/peft", "https://github.com/meta-llama/llama3" ]
29,261
0
Unknown
Unspecified
157.46
Test-Time Training (TTT) is an emerging paradigm that enables models to adapt their parameters during inference, improving performance on tasks such as few-shot learning, retrieval-augmented generation, and complex reasoning. However, this dynamic adaptation introduces new vulnerabilities that adversaries can exploit t...
[ 0.011111999861896038, 0.06760600209236145, 0.029092999175190926, 0.0629270002245903, 0.13667500019073486, 0.046891000121831894, 0.05154699832201004, 0.031189000234007835, -0.05886299908161163, 0.034752000123262405, 0.02966099977493286, 0.069193996489048, -0.011429999954998493, 0.0996259972...
[ -0.08491300046443939, 0.0051819998770952225, -0.03574499860405922, 0.04805000126361847, 0.07812900096178055, 0.037046000361442566, 0.04425700008869171, -0.05096599832177162, 0.021802999079227448, -0.041503001004457474, 0.0018179999897256494, -0.037735000252723694, 0.02657799981534481, 0.00...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Test-Time Training (TTT) is an emerging paradigm that enables models to adapt their parameters during inference, improving performance on tasks such as few-shot learning, retrieval-augmented generation, and complex reasoning.
However, this dynamic adaptation introduces new vulnerabilities that adversaries can exploit to jailbreak models.
We also show that TTT-induced overfitting can produce degenerate outputs that inflate ASR under standard judges, and propose a validity-aware evaluation to correct for this.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:34:25.814364
2606.13680v1
Learning to Reason by Analogy via Retrieval-Augmented Reinforcement Fine-Tuning
Enterprise RAG & Vector Search 2026
2026-06-11
[ "cs.CL", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Retrieval-augmented generation (RAG) has become a standard mechanism for grounding language models in external knowledge, yet conventional retrieval based on lexical or semantic similarity is poorly suited for complex reasoning tasks: a semantically
Zilin Xiao
7
[ "Zilin Xiao", "Qi Ma", "Chun-cheng Jason Chen", "Xintao Chen", "Avinash Atreya", "Hanjie Chen", "Vicente Ordonez" ]
[]
http://arxiv.org/abs/2606.13680v1
VERIFIED_LIVE
https://github.com/huggingface/open-r1
[ "https://github.com/huggingface/open-r1" ]
26,434
0
Unknown
Unspecified
157.4
Retrieval-augmented generation (RAG) has become a standard mechanism for grounding language models in external knowledge, yet conventional retrieval based on lexical or semantic similarity is poorly suited for complex reasoning tasks: a semantically similar problem may demand an entirely different solution strategy, wh...
[ 0.013167999684810638, -0.07575000077486038, 0.033642999827861786, 0.006399000063538551, 0.026605000719428062, 0.0248429998755455, 0.07665599882602692, 0.015568000264465809, 0.060398001223802567, 0.03625300154089928, 0.0012390000047162175, -0.010358000174164772, -0.0027280000504106283, 0.03...
[ -0.052069999277591705, -0.07309799641370773, 0.011141999624669552, 0.061253998428583145, 0.030786000192165375, 0.004255999810993671, 0.06720100343227386, 0.09306199848651886, -0.010758999735116959, -0.020752999931573868, -0.028578000143170357, -0.03660000115633011, 0.04209500178694725, 0.0...
Enterprise RAG & Vector Search 2026
git clone https://github.com/huggingface/open-r1
Retrieval-augmented generation (RAG) has become a standard mechanism for grounding language models in external knowledge, yet conventional retrieval based on lexical or semantic similarity is poorly suited for complex reasoning tasks: a semantically similar problem may demand an entirely different solution strategy, wh...
We propose Retrieval-Augmented Reinforcement Fine-Tuning (RA-RFT), a post-training framework that teaches language models to reason by analogy.
For example, it improves AIME 2025 average@32 accuracy by 7.1 and 2.8 points over GRPO for Qwen3-1.7B and Qwen3-4B respectively -- suggesting that reasoning-aware retrieval is a complementary axis of improvement and orthogonal to advances in reward design or training curricula.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:19.979110
2605.26533v1
A Hybrid Vision-Language Architecture for Automated Defect Reasoning and Report Generation in Industrial Inspection
Enterprise RAG & Vector Search 2026
2026-05-26
[ "cs.CV", "cs.AI", "cs.CL", "cs.LG" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Automated industrial inspection requires both precise defect localization and structured maintenance report generation; in current practice these tasks are handled separately, with linguistic interpretation left to human experts. This paper describes
Malikussaid
2
[ "Malikussaid", "Imad Gohar" ]
[]
http://arxiv.org/abs/2605.26533v1
VERIFIED_LIVE
https://github.com/HumanSignal/label-studio
[ "https://github.com/imadgohar/DTU-annotations", "https://github.com/HumanSignal/label-studio" ]
28,046
0
Unknown
Unspecified
157.25
Automated industrial inspection requires both precise defect localization and structured maintenance report generation; in current practice these tasks are handled separately, with linguistic interpretation left to human experts. This paper describes a decoupled, edge-deployable pipeline for wind turbine blade inspecti...
[ -0.018211999908089638, 0.012796999886631966, 0.02771800011396408, 0.005235999822616577, 0.04197400063276291, -0.018316000699996948, 0.04938200116157532, 0.09422799944877625, -0.05075399950146675, -0.029725000262260437, -0.0016540000215172768, -0.09142500162124634, 0.07077600061893463, 0.06...
[ -0.016443999484181404, 0.01562800072133541, 0.05127999931573868, -0.009855999611318111, 0.05273599922657013, -0.05628100037574768, 0.012435000389814377, 0.008346999995410442, -0.03632500022649765, -0.04780200123786926, -0.02505600079894066, -0.05379899963736534, 0.027935000136494637, 0.056...
Enterprise RAG & Vector Search 2026
git clone https://github.com/HumanSignal/label-studio
Automated industrial inspection requires both precise defect localization and structured maintenance report generation; in current practice these tasks are handled separately, with linguistic interpretation left to human experts.
This paper describes a decoupled, edge-deployable pipeline for wind turbine blade inspection built from three components that each handle a distinct sub-task.
The results show that purpose-built decoupled architecture with a small domain-specific training corpus outperforms a generalist end-to-end model on this structured generation task.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:39.098656
2605.08513v1
A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models
AI Security, Red Teaming & Defense 2026
2026-05-08
[ "cs.CL", "cs.AI", "cs.LG" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Safety alignment in language models operates through two mechanistically distinct systems: refusal neurons that gate whether harmful knowledge is expressed, and concept neurons that encode the harmful knowledge itself. By targeting a single neuron in
Hamid Kazemi
3
[ "Hamid Kazemi", "Atoosa Chegini", "Maria Safi" ]
[]
http://arxiv.org/abs/2605.08513v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/tatsu-lab/stanford_alpaca" ]
30,247
3,991
2024-07-17
Apache-2.0
157.17
Safety alignment in language models operates through two mechanistically distinct systems: refusal neurons that gate whether harmful knowledge is expressed, and concept neurons that encode the harmful knowledge itself. By targeting a single neuron in each system, we demonstrate both directions of failure -- bypassing s...
[ -0.014669000171124935, -0.10430999845266342, 0.013199999928474426, -0.06361699849367142, 0.008740999735891819, 0.08855000138282776, 0.016155000776052475, 0.01422599982470274, 0.12840400636196136, 0.024605000391602516, 0.030424000695347786, -0.029746999964118004, 0.013892999850213528, -0.00...
[ 0.03236199915409088, -0.08018600195646286, -0.021640000864863396, -0.03568200021982193, 0.034060999751091, 0.04475900158286095, 0.006136999931186438, 0.011481000110507011, 0.09388399869203568, 0.01786400005221367, 0.010254999622702599, -0.02006400004029274, 0.09043300151824951, 0.009066999...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Safety alignment in language models operates through two mechanistically distinct systems: refusal neurons that gate whether harmful knowledge is expressed, and concept neurons that encode the harmful knowledge itself.
By targeting a single neuron in each system, we demonstrate both directions of failure -- bypassing safety on explicit harmful requests via suppression, and inducing harmful content from innocent prompts via amplification -- across seven models spanning two families and 1.7B to 70B parameters, without any training or p...
By targeting a single neuron in each system, we demonstrate both directions of failure -- bypassing safety on explicit harmful requests via suppression, and inducing harmful content from innocent prompts via amplification -- across seven models spanning two families and 1.7B to 70B parameters, without any training or p...
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:39:59.551813
2605.04572v1
From Parameter Dynamics to Risk Scoring : Quantifying Sample-Level Safety Degradation in LLM Fine-tuning
AI Security, Red Teaming & Defense 2026
2026-05-06
[ "cs.AI", "cs.LG" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Safety alignment of Large Language Models (LLMs) is extremely fragile, as fine-tuning on a small number of benign samples can erase safety behaviors learned from millions of preference examples. Existing studies attempt to explain this phenomenon by
Xiao Wang
7
[ "Xiao Wang", "Yifei Zhang", "YongKang Liu", "Xiaocui Yang", "Zihan Wang", "Shi Feng", "Daling Wang" ]
[]
http://arxiv.org/abs/2605.04572v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/tatsu-lab/stanford_alpaca" ]
30,247
3,991
2024-07-17
Apache-2.0
157.07
Safety alignment of Large Language Models (LLMs) is extremely fragile, as fine-tuning on a small number of benign samples can erase safety behaviors learned from millions of preference examples. Existing studies attempt to explain this phenomenon by comparing parameters and hidden states before and after fine-tuning, b...
[ 0.07288400083780289, -0.0591449998319149, -0.0277670007199049, -0.008065000176429749, 0.1289840042591095, 0.010898999869823456, -0.01676500029861927, 0.09515900164842606, -0.028113000094890594, 0.008558000437915325, -0.03061399981379509, 0.010517000220716, 0.03875400125980377, -0.035884998...
[ 0.011122000403702259, -0.11605499684810638, -0.002355000004172325, 0.020914999768137932, 0.11883500218391418, 0.04978900030255318, -0.0040790000930428505, 0.035757001489400864, 0.027111999690532684, -0.016681000590324402, -0.01612900011241436, -0.009518999606370926, 0.05670800060033798, -0...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Safety alignment of Large Language Models (LLMs) is extremely fragile, as fine-tuning on a small number of benign samples can erase safety behaviors learned from millions of preference examples.
In this paper, we uncover a critical mechanism underlying safety degradation by analyzing parameter dynamics, where benign fine-tuning causes parameters to cumulatively drift toward danger-aligned directions, progressively undermining the model's safety.
Extensive experiments across multiple models and datasets demonstrate that SQSD effectively quantifies sample-level fine-tuning risks and exhibits strong transferability across model architectures, parameter scales, and parameter-efficient methods.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:41:51.393326
2605.14034v1
From Descriptive to Prescriptive: Uncover the Social Value Alignment of LLM-based Agents
Enterprise RAG & Vector Search 2026
2026-05-13
[ "cs.AI", "cs.CL", "cs.CY" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Wide applications of LLM-based agents require strong alignment with human social values. However, current works still exhibit deficiencies in self-cognition and dilemma decision, as well as self-emotions. To remedy this, we propose a novel value-base
Jinxian Qu
4
[ "Jinxian Qu", "Qingqing Gu", "Teng Chen", "Luo Ji" ]
[]
http://arxiv.org/abs/2605.14034v1
VERIFIED_LIVE
https://github.com/meta-llama/llama3
[ "https://github.com/meta-llama/llama3" ]
29,261
0
Unknown
Unspecified
157.06
Wide applications of LLM-based agents require strong alignment with human social values. However, current works still exhibit deficiencies in self-cognition and dilemma decision, as well as self-emotions. To remedy this, we propose a novel value-based framework that employs GraphRAG to convert principles into value-bas...
[ 0.055358000099658966, -0.026434000581502914, -0.046521998941898346, -0.053697001188993454, -0.0018990000244230032, 0.06202400103211403, -0.007375000044703484, -0.03073200024664402, 0.012810000218451023, -0.007521999999880791, 0.04539500176906586, -0.036591000854969025, 0.06560099869966507, ...
[ 0.035041000694036484, 0.04325199872255325, -0.044697001576423645, -0.033663999289274216, -0.010974000208079815, -0.044895000755786896, 0.045412998646497726, 0.08156699687242508, 0.029308000579476357, -0.005262000020593405, -0.014531999826431274, -0.06123200058937073, 0.03341799974441528, 0...
Enterprise RAG & Vector Search 2026
git clone https://github.com/meta-llama/llama3
Wide applications of LLM-based agents require strong alignment with human social values.
To remedy this, we propose a novel value-based framework that employs GraphRAG to convert principles into value-based instructions and steer the agent to behave as expected by retrieving the suitable instruction upon a specific conversation context.
Our method provides a basis for the emergence of self-emotion in AI systems.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:44.779922
2605.22880v1
How Far Will They Go? Red-Teaming Online Influence with Large Language Models
AI Security, Red Teaming & Defense 2026
2026-05-20
[ "cs.CL", "cs.AI", "cs.CY" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
As large language model (LLM)-based agents increasingly participate in online discourse, red-teaming their capacity to support political influence campaigns is critical for information integrity. In pursuit of this goal, we focus on locally deployed
Daniel C. Ruiz
5
[ "Daniel C. Ruiz", "Anna Serbina", "Ashwin Rao", "Emilio Ferrara", "Luca Luceri" ]
[]
http://arxiv.org/abs/2605.22880v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/elder-plinius/OBLITERATUS", "https://github.com/SIGNALS-Lab/llm-overton-external", "https://github.com/p-e-w/heretic" ]
27,374
0
Unknown
Unspecified
156.68
As large language model (LLM)-based agents increasingly participate in online discourse, red-teaming their capacity to support political influence campaigns is critical for information integrity. In pursuit of this goal, we focus on locally deployed open-source LLMs, as opposed to frontier API-only models, given their ...
[ 0.013194999657571316, -0.14253300428390503, -0.02153399959206581, 0.016819000244140625, 0.021012000739574432, 0.0826219990849495, -0.002220999915152788, 0.03101399913430214, 0.05597500130534172, -0.10107699781656265, -0.004255000036209822, 0.032524000853300095, 0.03909499943256378, -0.0238...
[ -0.020419999957084656, -0.12723299860954285, -0.027861999347805977, -0.009479999542236328, 0.1393740028142929, 0.012502999976277351, -0.02150299958884716, -0.05501899868249893, 0.061260998249053955, -0.004476000089198351, -0.03494799882173538, 0.034088000655174255, 0.05903699994087219, -0....
AI Security, Red Teaming & Defense 2026
No code repository attached.
As large language model (LLM)-based agents increasingly participate in online discourse, red-teaming their capacity to support political influence campaigns is critical for information integrity.
We introduce an empirical red-teaming framework for measuring LLM Overton Windows (OWs), defined as the range of political opinions a model can reliably express on controversial topics, and for quantifying how simple natural-language jailbreaks expand that range.
Taken together, our results establish a practical framework for auditing the political steerability of open-source LLMs and for helping future researchers design stronger countermeasures against LLM-enabled influence campaigns.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:34:42.550833
2605.01899v1
Disentangling Intent from Role: Adversarial Self-Play for Persona-Invariant Safety Alignment
AI Security, Red Teaming & Defense 2026
2026-05-03
[ "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
The growing capabilities of large language models (LLMs) have driven their widespread deployment across diverse domains, even in potentially high-risk scenarios. Despite advances in safety alignment techniques, current models remain vulnerable to eme
Jiajia Li
6
[ "Jiajia Li", "Xiaoyu Wen", "Zhongtian Ma", "Shuyue Hu", "Qiaosheng Zhang", "Zhen Wang" ]
[]
http://arxiv.org/abs/2605.01899v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/huggingface/trl", "https://github.com/JiajiaLi-1130/PIA", "https://github.com/meta-llama/llama3" ]
29,261
0
Unknown
Unspecified
156.56
The growing capabilities of large language models (LLMs) have driven their widespread deployment across diverse domains, even in potentially high-risk scenarios. Despite advances in safety alignment techniques, current models remain vulnerable to emerging persona-based jailbreak attacks. Existing research on persona-ba...
[ -0.05314299836754799, 0.0011309999972581863, -0.01770699955523014, -0.11141300201416016, 0.0486689992249012, 0.04452599957585335, 0.07188300043344498, -0.00698600010946393, 0.03431899845600128, 0.005111000034958124, 0.028020000085234642, -0.05738599970936775, -0.021988999098539352, 0.01718...
[ -0.10785699635744095, -0.03658200055360794, -0.0010499999625608325, -0.09129299968481064, 0.008004000410437584, 0.06398700177669525, 0.03237000107765198, -0.07123500108718872, 0.05369099974632263, -0.02423899993300438, 0.06308899819850922, -0.06141899898648262, 0.031831998378038406, -0.020...
AI Security, Red Teaming & Defense 2026
No code repository attached.
The growing capabilities of large language models (LLMs) have driven their widespread deployment across diverse domains, even in potentially high-risk scenarios.
Despite advances in safety alignment techniques, current models remain vulnerable to emerging persona-based jailbreak attacks.
Meanwhile, the PICL defense method significantly reduces the Attack Success Rate (ASR) while preserving the model's general capability, thereby validating the superiority and robustness of this alignment paradigm.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:42:47.593165
2606.28781v1
HyphaeDB: A Living Knowledge Topology for Agent-First Memory
Enterprise RAG & Vector Search 2026
2026-06-27
[ "cs.AI", "cs.MA" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Every existing vector database and agent memory framework treats memory as passive storage that agents query explicitly. No system propagates knowledge between agents through the memory layer itself. We introduce HyphaeDB, an agent-native memory infr
Krishna Halaharvi
1
[ "Krishna Halaharvi" ]
[]
http://arxiv.org/abs/2606.28781v1
VERIFIED_LIVE
https://github.com/pgvector/pgvector
[ "https://github.com/pgvector/pgvector" ]
22,615
0
Unknown
Unspecified
156.51
Every existing vector database and agent memory framework treats memory as passive storage that agents query explicitly. No system propagates knowledge between agents through the memory layer itself. We introduce HyphaeDB, an agent-native memory infrastructure that reinterprets the Hierarchical Navigable Small World (H...
[ 0.034380000084638596, -0.050932999700307846, -0.11455799639225006, 0.007969999685883522, -0.004912999924272299, 0.009425999596714973, 0.03700299933552742, -0.009286999702453613, 0.0014220000011846423, 0.002351999981328845, 0.012772999703884125, -0.01786700077354908, 0.07069700211286545, 0....
[ 0.06371799856424332, -0.02353000082075596, -0.09397699683904648, 0.04976600036025047, -0.0000019999999949504854, -0.018178999423980713, -0.02912200056016445, -0.03445800021290779, -0.07987699657678604, 0.04674199968576431, -0.015425000339746475, 0.016397999599575996, 0.08957900106906891, 0...
Enterprise RAG & Vector Search 2026
git clone https://github.com/pgvector/pgvector
Every existing vector database and agent memory framework treats memory as passive storage that agents query explicitly.
No system propagates knowledge between agents through the memory layer itself.
HyphaeDB represents, to our knowledge, the first system to combine navigable small world topology with gossip-based knowledge propagation for multi-agent coordination.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:05.657287
2605.12028v1
Caraman at SemEval-2026 Task 8: Three-Stage Multi-Turn Retrieval with Query Rewriting, Hybrid Search, and Cross-Encoder Reranking
Enterprise RAG & Vector Search 2026
2026-05-12
[ "cs.CL", "cs.IR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
We describe our system for SemEval-2026 Task 8 (MTRAGEval), participating in Task A (Retrieval) across four English-language domains. Our approach employs a three-stage pipeline: (1) query rewriting via a LoRA-fine-tuned Qwen 2.5 7B model that transf
David-Maximilian Caraman
2
[ "David-Maximilian Caraman", "Gheorghe Cosmin Silaghi" ]
[]
http://arxiv.org/abs/2605.12028v1
VERIFIED_LIVE
https://github.com/ml-explore/mlx
[ "https://github.com/ml-explore/mlx", "https://github.com/davidcaraman/semeval2026-mtrag-retrieval" ]
27,931
0
Unknown
Unspecified
156.5
We describe our system for SemEval-2026 Task 8 (MTRAGEval), participating in Task A (Retrieval) across four English-language domains. Our approach employs a three-stage pipeline: (1) query rewriting via a LoRA-fine-tuned Qwen 2.5 7B model that transforms context-dependent follow-up questions into standalone queries, (2...
[ -0.04867099970579147, -0.018326999619603157, 0.05579699948430061, -0.009211000055074692, -0.0439470000565052, 0.0812930017709732, -0.04979399964213371, 0.018604999408125877, -0.028279999271035194, -0.046140000224113464, -0.012849999591708183, -0.03948799893260002, 0.024163000285625458, -0....
[ -0.05929800122976303, -0.08225999772548676, 0.02496900036931038, 0.04669800028204918, -0.017681000754237175, 0.020291000604629517, 0.009031999856233597, 0.00582600012421608, -0.012360000051558018, -0.024150000885128975, -0.0620650015771389, -0.08347199857234955, 0.11155900359153748, -0.007...
Enterprise RAG & Vector Search 2026
git clone https://github.com/ml-explore/mlx
We describe our system for SemEval-2026 Task 8 (MTRAGEval), participating in Task A (Retrieval) across four English-language domains.
On the official test set, the system achieves nDCG@5 of 0.531, ranking 8th out of 38 participating systems and 10.7% above the organizer baseline.
On the official test set, the system achieves nDCG@5 of 0.531, ranking 8th out of 38 participating systems and 10.7% above the organizer baseline.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:52.803722
2606.30755v1
Understanding and Evaluating Claw-like Agent Security Through a Computer-Systems Lens
AI Security, Red Teaming & Defense 2026
2026-06-29
[ "cs.CR", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Claw-like AI agents (e.g., OpenClaw) are always-on processes with persistent access to credentials, files, tools, and external services. They take on system-level responsibilities -- installing packages, maintaining state, scheduling subtasks, and me
Peizhi Niu
17
[ "Peizhi Niu", "Wenjie Qu", "Shangding Gu", "Tianneng Shi", "Yuankai Li", "Ahmad Tawaha", "Hend Alzahrani", "Vincent Siu", "Boyi Li", "Chenguang Wang", "Jiaheng Zhang", "Basel Alomair", "Ming Jin", "Muhao Chen", "Chi Wang", "Costas Spanos", "Dawn Song" ]
[]
http://arxiv.org/abs/2606.30755v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/jgamblin/OpenClawCVEs", "https://github.com/NVIDIA/NemoClaw", "https://github.com/sunblaze-ucb/SafeClawArena" ]
22,143
3,017
2026-07-06
MIT
156.38
Claw-like AI agents (e.g., OpenClaw) are always-on processes with persistent access to credentials, files, tools, and external services. They take on system-level responsibilities -- installing packages, maintaining state, scheduling subtasks, and mediating I/O -- making security failures far more severe than in other ...
[ -0.09167800098657608, 0.054315000772476196, -0.05329100042581558, -0.09411100298166275, -0.023656999692320824, -0.04439299926161766, 0.07385499775409698, 0.012772000394761562, 0.0287299994379282, 0.06137600168585777, -0.003011999884620309, 0.004383999854326248, 0.069923996925354, 0.0070870...
[ -0.10029099881649017, -0.017105000093579292, -0.027295000851154327, -0.008740999735891819, 0.026264000684022903, -0.07134199887514114, 0.0030769999139010906, 0.04395399987697601, 0.007797999773174524, -0.023517999798059464, -0.07204800099134445, -0.02687999978661537, 0.062182001769542694, ...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Claw-like AI agents (e.g., OpenClaw) are always-on processes with persistent access to credentials, files, tools, and external services.
They take on system-level responsibilities -- installing packages, maintaining state, scheduling subtasks, and mediating I/O -- making security failures far more severe than in other agents.
These results expose the inadequacy of current defenses and suggest directions for future hardening.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:22:47.694676
2605.14362v1
Correctness-Aware Repository Filtering Under Maximum Effective Context Window Constraints
Enterprise RAG & Vector Search 2026
2026-05-14
[ "cs.SE", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Context window efficiency is a practical constraint in large language model (LLM)-based developer tools. Paulsen [12] shows that all tested models degrade in accuracy well before their advertised context limits the Maximum Effective Context Window (M
Shweta Mishra
1
[ "Shweta Mishra" ]
[]
http://arxiv.org/abs/2605.14362v1
VERIFIED_LIVE
https://github.com/tree-sitter/tree-sitter
[ "https://github.com/tree-sitter/tree-sitter", "https://github.com/features/copilot", "https://github.com/openai/tiktoken" ]
26,631
0
Unknown
Unspecified
156.09
Context window efficiency is a practical constraint in large language model (LLM)-based developer tools. Paulsen [12] shows that all tested models degrade in accuracy well before their advertised context limits the Maximum Effective Context Window (MECW) which makes context construction a quality problem, not just a co...
[ -0.030067000538110733, -0.028169000521302223, -0.02084900066256523, 0.019655000418424606, 0.04028800129890442, -0.06794799864292145, 0.11854500323534012, -0.08564099669456482, -0.01750900037586689, 0.03698499873280525, -0.040518999099731445, -0.05018800124526024, 0.06213900074362755, 0.028...
[ -0.027588000521063805, 0.02056100033223629, -0.0157919991761446, 0.05110999941825867, 0.08715000003576279, -0.11373399943113327, 0.044860001653432846, 0.010110000148415565, -0.005044000223278999, -0.00216599996201694, -0.0664369985461235, -0.044822998344898224, 0.012190000154078007, -0.016...
Enterprise RAG & Vector Search 2026
git clone https://github.com/tree-sitter/tree-sitter
Context window efficiency is a practical constraint in large language model (LLM)-based developer tools.
We present a correctness-aware context hygiene framework: a pre-execution, size-based heuristic filter that intercepts repository scans before tokenization, using only OS-level stat() metadata with sub-millisecond overhead.
Across 10 real open-source repositories (22,046 files, 5 languages), the proposed SizeFilter at θ=1 MB achieves 79.6% (\pm13.2%) mean token reduction at 0.30 ms overhead: the HybridFilter achieves 89.3% (\pm9.0%) the lowest variance of any filter evaluated.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:49.791134
2606.19803v1
Policy-aware Vector Search: A Vision for Fine Grained Access Control in Vector Databases
Enterprise RAG & Vector Search 2026
2026-06-18
[ "cs.DB", "cs.AI", "cs.LG" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Vector databases are increasingly used in security sensitive contexts with Retrieval Augmented Generation and organizational AI pipelines; however, their security capabilities remain limited. Specifically, Fine-grained Access Control (FGAC) which is
Lakshmi Sahithi Yalamarthi
2
[ "Lakshmi Sahithi Yalamarthi", "Primal Pappachan" ]
[]
http://arxiv.org/abs/2606.19803v1
VERIFIED_LIVE
https://github.com/pgvector/pgvector
[ "https://github.com/pgvector/pgvector" ]
22,615
0
Unknown
Unspecified
156.06
Vector databases are increasingly used in security sensitive contexts with Retrieval Augmented Generation and organizational AI pipelines; however, their security capabilities remain limited. Specifically, Fine-grained Access Control (FGAC) which is required to ensure that data access adheres to user-specific policies ...
[ -0.002879000036045909, 0.025366999208927155, -0.04785899817943573, 0.02339700050652027, 0.012915000319480896, -0.010087000206112862, 0.028067000210285187, -0.06968200206756592, -0.0628110021352768, 0.030772000551223755, -0.058625999838113785, 0.050140999257564545, 0.06613700091838837, -0.0...
[ -0.01618500053882599, 0.0287260003387928, -0.09078499674797058, 0.05544000118970871, -0.015406999737024307, -0.0012710000155493617, 0.04632100090384483, -0.05399100109934807, -0.04237699881196022, 0.007017000112682581, -0.07195699959993362, 0.034749001264572144, 0.03850099816918373, -0.021...
Enterprise RAG & Vector Search 2026
git clone https://github.com/pgvector/pgvector
Vector databases are increasingly used in security sensitive contexts with Retrieval Augmented Generation and organizational AI pipelines; however, their security capabilities remain limited.
In this paper, we present a vision for Policy-aware Vector Search by formalizing the FGAC policy model in vector databases as well as the enforcement problem.
Unlike relational databases, vector databases combine structured and unstructured attributes to provide semantic, approximate query results, which complicates FGAC implementation.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:10.485811
2606.26377v1
Verifying Intent and Harm: A Unified Defense Against LLM-Generated Threats
AI Security, Red Teaming & Defense 2026
2026-06-24
[ "cs.CR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Large language models (LLMs) are increasingly deployed in interactive applications, yet they remain vulnerable to adversarial interactions that induce harmful, deceptive, or policy-violating outputs. Existing defenses typically analyze either user pr
Poojitha Thota
5
[ "Poojitha Thota", "Yun Lei", "Santhosh Thangaraj", "Siddhartha Reddy Jonnalagadda", "Shirin Nilizadeh" ]
[]
http://arxiv.org/abs/2606.26377v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/openai/swarm" ]
21,903
2,329
2026-04-15
MIT
156.01
Large language models (LLMs) are increasingly deployed in interactive applications, yet they remain vulnerable to adversarial interactions that induce harmful, deceptive, or policy-violating outputs. Existing defenses typically analyze either user prompts or generated outputs, but not both. However, many real-world att...
[ -0.016454000025987625, 0.043088000267744064, -0.0035689999349415302, -0.041482001543045044, 0.11217799782752991, 0.040748998522758484, 0.036180999130010605, -0.030570000410079956, 0.02150299958884716, 0.045921001583337784, 0.02025200054049492, -0.0225600004196167, 0.11708000302314758, 0.00...
[ -0.042017001658678055, -0.004098999779671431, -0.0038509999867528677, -0.026972999796271324, 0.09476400166749954, 0.02186499908566475, -0.007091000210493803, -0.028842000290751457, 0.017621999606490135, -0.02194399945437908, -0.06871400028467178, -0.10294400155544281, 0.06780599802732468, ...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Large language models (LLMs) are increasingly deployed in interactive applications, yet they remain vulnerable to adversarial interactions that induce harmful, deceptive, or policy-violating outputs.
Existing defenses typically analyze either user prompts or generated outputs, but not both.
Our results suggest that prompt-response verification provides a practical foundation for securing LLM applications against evolving adversarial threats.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:23:29.852249
2605.17310v1
Attention Hijacking: Response Manipulation Across Queries in Vision-Language Models
AI Security, Red Teaming & Defense 2026
2026-05-17
[ "cs.CV", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Existing adversarial attacks on vision-language models (VLMs) can steer model outputs toward attacker-specified target responses, but their effectiveness often degrades when the same perturbed input is paired with different textual queries. This pape
Zhiqiang Wang
7
[ "Zhiqiang Wang", "Dongrui Liu", "Yan Li", "Zonghao Ying", "Wei Xue", "Wenhan Luo", "Yike Guo" ]
[]
http://arxiv.org/abs/2605.17310v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/black-forest-labs/flux" ]
25,898
0
Unknown
Unspecified
155.93
Existing adversarial attacks on vision-language models (VLMs) can steer model outputs toward attacker-specified target responses, but their effectiveness often degrades when the same perturbed input is paired with different textual queries. This paper studies cross-query response manipulation, where a single adversaria...
[ 0.012470999732613564, -0.03779999911785126, 0.013667000457644463, 0.049320001155138016, 0.004011999815702438, 0.01105399988591671, 0.1132429987192154, 0.009171999990940094, 0.08737500011920929, -0.03422300145030022, 0.019812000915408134, -0.02541700005531311, 0.07359900325536728, -0.018440...
[ -0.012895000167191029, -0.047311000525951385, -0.00005999999848427251, 0.08200599998235703, 0.07910799980163574, 0.04145300015807152, 0.05006299912929535, -0.036052998155355453, 0.04268399998545647, -0.05639300122857094, -0.01945899985730648, -0.05842600017786026, 0.08030100166797638, -0.0...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Existing adversarial attacks on vision-language models (VLMs) can steer model outputs toward attacker-specified target responses, but their effectiveness often degrades when the same perturbed input is paired with different textual queries.
Motivated by the observation, we propose \textbf{Attention Hijacking}, a novel adversarial attack that explicitly steers internal attention distributions toward a persistent image-dominant pattern.
Extensive experiments on widely used VLMs show that Attention Hijacking substantially improves cross-query transferability across diverse target responses and unseen queries.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:36:24.118631
2605.23994v1
RAW: Robust Avatar Watermarking -- Benchmarking and Baseline
AI Security, Red Teaming & Defense 2026
2026-05-17
[ "cs.CV", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Digital avatar watermarking presents unique challenges: avatars are routinely post-processed with background replacement, reframing, and format conversion before deployment. We introduce \textbf{RAW} (Robust Avatar Watermarking), a benchmark comprisi
Jack Parry
3
[ "Jack Parry", "Jack Saunders", "Vinay Namboodiri" ]
[]
http://arxiv.org/abs/2605.23994v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/danielgatis/rembg" ]
24,240
0
Unknown
Unspecified
155.21
Digital avatar watermarking presents unique challenges: avatars are routinely post-processed with background replacement, reframing, and format conversion before deployment. We introduce \textbf{RAW} (Robust Avatar Watermarking), a benchmark comprising 50 synthetic avatar videos from 5 commercial providers and 6 attack...
[ -0.014543999917805195, 0.019331999123096466, 0.025617999956011772, -0.04427200183272362, 0.1042179986834526, 0.005359000060707331, 0.036430999636650085, -0.04125399887561798, 0.003763000015169382, 0.011256000027060509, -0.046275001019239426, -0.0882050022482872, 0.04454600065946579, 0.0335...
[ -0.08189599961042404, 0.03691599890589714, 0.06166600063443184, -0.02638000063598156, 0.1372690051794052, -0.03336599841713905, 0.06966199725866318, -0.08419399708509445, 0.005251999944448471, 0.0008570000063627958, -0.032621998339891434, -0.0004400000034365803, 0.02835099957883358, 0.0234...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Digital avatar watermarking presents unique challenges: avatars are routinely post-processed with background replacement, reframing, and format conversion before deployment.
We introduce \textbf{RAW} (Robust Avatar Watermarking), a benchmark comprising 50 synthetic avatar videos from 5 commercial providers and 6 attacks simulating real-world avatar workflows.
WALT achieves the highest robustness to zoom attacks (92.4\%) while maintaining strong performance on background removal (95.6\%).
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:36:05.139138
2606.00734v1
EMA: Approximate Nearest Neighbor Search with General Attribute Filtering and Dynamic Updates
Enterprise RAG & Vector Search 2026
2026-05-30
[ "cs.DB" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Filtering Approximate Nearest Neighbor (FANN) search is a critical and emerging task for strengthening the query capability of vector databases, supporting applications such as recommendation systems, retrieval-augmented generation (RAG), and agent m
Mocheng Li
4
[ "Mocheng Li", "Baotong Lu", "James Cheng", "Chenhao Ma" ]
[]
http://arxiv.org/abs/2606.00734v1
VERIFIED_LIVE
https://github.com/pgvector/pgvector
[ "https://github.com/lmccccc/EMA", "https://github.com/pgvector/pgvector" ]
22,615
0
Unknown
Unspecified
155.11
Filtering Approximate Nearest Neighbor (FANN) search is a critical and emerging task for strengthening the query capability of vector databases, supporting applications such as recommendation systems, retrieval-augmented generation (RAG), and agent memory. However, most existing methods are limited to range or label fi...
[ -0.03401999920606613, 0.03964700177311897, 0.05535599961876869, -0.07478500157594681, 0.06884600222110748, -0.031289998441934586, 0.005404000170528889, -0.04368700087070465, 0.0072639998979866505, -0.027998000383377075, 0.05425399914383888, -0.02305999957025051, 0.07788500189781189, -0.008...
[ 0.033723000437021255, 0.03720900043845177, 0.020036999136209488, -0.007703999988734722, 0.09440899640321732, 0.06118199974298477, 0.03581999987363815, -0.04975200071930885, -0.03452799841761589, -0.04297000169754028, 0.001820000004954636, 0.013124999590218067, 0.04792499914765358, -0.00558...
Enterprise RAG & Vector Search 2026
git clone https://github.com/pgvector/pgvector
Filtering Approximate Nearest Neighbor (FANN) search is a critical and emerging task for strengthening the query capability of vector databases, supporting applications such as recommendation systems, retrieval-augmented generation (RAG), and agent memory.
However, most existing methods are limited to range or label filtering, often incurring unacceptable index construction time and memory overhead.
Extensive experiments demonstrate that EMA achieves 1.68x--12.25x speedup over state-of-the-art general filtering ANN methods across diverse workloads.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:31.109459
2606.00809v2
NBQ: Next-Best-Question for Dynamic Profiling
Enterprise RAG & Vector Search 2026
2026-05-30
[ "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Many real-world conversational settings for knowledge discovery, including podcasts, hiring screens, and marketplaces, require a purpose-driven understanding of a person. We study the Next-Best-Question (NBQ) problem: at each turn, an interviewer sho
Yimin Shi
4
[ "Yimin Shi", "Clarice Wang", "Haixun Wang", "Xiaokui Xiao" ]
[]
http://arxiv.org/abs/2606.00809v2
VERIFIED_LIVE
https://github.com/pgvector/pgvector
[ "https://github.com/Hanc1999/Next-Best-Question", "https://github.com/baidu/puck", "https://github.com/pgvector/pgvector" ]
22,615
0
Unknown
Unspecified
155.11
Many real-world conversational settings for knowledge discovery, including podcasts, hiring screens, and marketplaces, require a purpose-driven understanding of a person. We study the Next-Best-Question (NBQ) problem: at each turn, an interviewer should ask the question with the highest expected information gain given ...
[ -0.02353299967944622, 0.014228999614715576, -0.049490999430418015, 0.04124699905514717, 0.008542000316083431, 0.06975200027227402, 0.07272399961948395, 0.0556659996509552, -0.11989899724721909, -0.08524800091981888, -0.062150001525878906, -0.08296799659729004, -0.04029399901628494, 0.04621...
[ -0.07857300341129303, -0.005745999980717897, -0.009212000295519829, 0.006703999824821949, 0.003017999930307269, 0.012551999650895596, 0.06673400104045868, 0.027070000767707825, -0.01905299909412861, -0.050193000584840775, -0.12042500078678131, -0.055201999843120575, -0.0040480000898242, 0....
Enterprise RAG & Vector Search 2026
git clone https://github.com/pgvector/pgvector
Many real-world conversational settings for knowledge discovery, including podcasts, hiring screens, and marketplaces, require a purpose-driven understanding of a person.
We propose NBQ, a plug-and-play framework that seeds a diverse pool of candidate questions, maintains a compact and continuously updated user state, adaptively selects the next question within a turn budget, and distills the resulting free-form dialogue into a structured vector-based user profile.
Experiments show that NBQ improves user profiling quality by up to 13.6% and 14.0% in AC@T and AR@T, respectively, while QuickMatch accelerates retrieval by up to 22.9x with recall up to 0.989.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:31.664192
2605.17561v2
Automated Root-Cause Subclassification and No-Code Fix Generation for Invalid Bug Reports
Enterprise RAG & Vector Search 2026
2026-05-17
[ "cs.SE", "cs.AI", "cs.MA" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Issues faced when using software are reported in the form of bug reports. However, many bug reports are invalid, meaning they do not require code changes, and are resolved with a no-code fix. Manually determining the root cause of the invalid bug rep
Mahmut Furkan Gon
4
[ "Mahmut Furkan Gon", "Emre Dinc", "Tevfik Emre Sungur", "Eray Tuzun" ]
[]
http://arxiv.org/abs/2605.17561v2
VERIFIED_LIVE
https://github.com/brave/brave-browser
[ "https://github.com/frog1014/lazy_chrome", "https://github.com/brave/brave-core", "https://github.com/brave/brave-browser", "https://github.com/features/issues" ]
23,335
0
Unknown
Unspecified
154.8
Issues faced when using software are reported in the form of bug reports. However, many bug reports are invalid, meaning they do not require code changes, and are resolved with a no-code fix. Manually determining the root cause of the invalid bug reports and providing actionable resolutions by the customer support caus...
[ -0.11087799817323685, 0.0013340000296011567, 0.013949000276625156, 0.03616899996995926, 0.02692900039255619, -0.085548996925354, -0.08419600129127502, 0.05081700161099434, -0.14091700315475464, 0.01860100030899048, -0.03333200141787529, -0.03577800095081329, 0.030702000483870506, -0.014747...
[ -0.07722099870443344, -0.03826599940657616, -0.004939999897032976, 0.027646999806165695, 0.037797000259160995, -0.08927000313997269, -0.08471699804067612, 0.013782000169157982, -0.1293260008096695, -0.003100000089034438, -0.060277000069618225, -0.020552000030875206, 0.03854300081729889, -0...
Enterprise RAG & Vector Search 2026
git clone https://github.com/brave/brave-browser
Issues faced when using software are reported in the form of bug reports.
Our goal is to introduce a standardized taxonomy for root-cause oriented invalid bug report subclassification, and perform experiments to test the accuracy of various approaches on invalid subclassification and no-code fix generation.
For no-code fix generation, agentic web search achieves the highest overall Judge LLM success rate at 68.9%, compared to 64.4% for RAG applications and 64.9% for vanilla LLMs, with subclass-level peaks of 87.4% for Working as Designed and 72.2% for Question.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:48.409572
2606.02822v1
Which Defense Closes Which Threat? Attributing OWASP-LLM-Top-10 Coverage and Its Brittleness Under Paraphrasing
AI Security, Red Teaming & Defense 2026
2026-06-01
[ "cs.CR", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Production LLM applications stack several defense families -- refusal-phrase filters, token-budget controls, model allowlists, rate limits, tool-registry authentication -- yet existing breach-and-attack-simulation (BAS) benchmarks report a single agg
Alexandre Cristovão Maiorano
1
[ "Alexandre Cristovão Maiorano" ]
[]
http://arxiv.org/abs/2606.02822v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/mitre-engenuity/ctid-attack-evaluations", "https://github.com/alemaiorano/llm-defense-lattice", "https://github.com/elder-plinius/L1B3RT4S" ]
20,907
0
Unknown
Unspecified
154.36
Production LLM applications stack several defense families -- refusal-phrase filters, token-budget controls, model allowlists, rate limits, tool-registry authentication -- yet existing breach-and-attack-simulation (BAS) benchmarks report a single aggregate coverage number, hiding which family closes which threat. We me...
[ -0.023019999265670776, 0.03655900061130524, -0.044461000710725784, -0.00776899978518486, 0.12564900517463684, 0.08692800253629684, 0.024709999561309814, 0.04232899844646454, 0.014461000449955463, 0.0682239979505539, -0.02097499929368496, -0.027535999193787575, 0.08262500166893005, 0.036010...
[ -0.06131500005722046, -0.04004799947142601, -0.0064570000395178795, -0.0021299999207258224, 0.0825980007648468, 0.005371999926865101, -0.01553099974989891, 0.012870999984443188, 0.04613799974322319, 0.025523999705910683, -0.002856000093743205, -0.04997999966144562, 0.10014200210571289, -0....
AI Security, Red Teaming & Defense 2026
No code repository attached.
Production LLM applications stack several defense families -- refusal-phrase filters, token-budget controls, model allowlists, rate limits, tool-registry authentication -- yet existing breach-and-attack-simulation (BAS) benchmarks report a single aggregate coverage number, hiding which family closes which threat.
We add four OWASP-LLM-Top-10-aware agents to a 21-agent baseline scanner and target a lattice of four synthetic LLM endpoints: $L_0$ (no defenses), $L_1$ (refusal-only), $L_2$ (budget-only), and $L_3$ (full stack).
A refusal whitelist that clears a static benchmark can be defeated by an LLM-driven paraphraser without changing attack intent; a budget control resists the same mutation.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:30:09.785189
2606.18801v1
SHIFT: Semantic Harmonization via Index-side Feature Transformation for Multilingual Information Retrieval
Enterprise RAG & Vector Search 2026
2026-06-17
[ "cs.IR", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
With the rapid expansion of massive multilingual corpora, Multilingual Information Retrieval (MLIR) has emerged as a critical technology for global information access. MLIR enables users to retrieve semantically relevant documents from multilingual t
Youngjoon Jang
4
[ "Youngjoon Jang", "Seongtae Hong", "Hyeonseok Moon", "Heuiseok Lim" ]
[]
http://arxiv.org/abs/2606.18801v1
VERIFIED_LIVE
https://github.com/huggingface/sentence-transformers
[ "https://github.com/huggingface/sentence-transformers" ]
18,998
0
Unknown
Unspecified
154.12
With the rapid expansion of massive multilingual corpora, Multilingual Information Retrieval (MLIR) has emerged as a critical technology for global information access. MLIR enables users to retrieve semantically relevant documents from multilingual text collections using a single-language query. However, recent multili...
[ -0.029160000383853912, 0.008642000146210194, 0.05789799988269806, -0.00407399982213974, 0.001988000003620982, 0.0549280010163784, -0.005803999956697226, -0.0007089999853633344, 0.017826000228524208, -0.05404699966311455, -0.001019999966956675, -0.027053000405430794, 0.012442000210285187, 0...
[ -0.008984999731183052, -0.08144699782133102, 0.017148999497294426, 0.0045449999161064625, 0.038986001163721085, 0.03282900154590607, 0.015402999706566334, 0.03635900095105171, 0.05583300068974495, -0.029756000265479088, 0.013783000409603119, 0.007903999648988247, 0.015811000019311905, 0.02...
Enterprise RAG & Vector Search 2026
git clone https://github.com/huggingface/sentence-transformers
With the rapid expansion of massive multilingual corpora, Multilingual Information Retrieval (MLIR) has emerged as a critical technology for global information access.
To address this issue, we propose SHIFT, a training-free method applicable in the indexing stage.
This leads to severe language bias, where top-ranked results are dominated by documents of specific languages, even when documents in other languages contain more semantically relevant information.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:14.906032
2606.30153v1
The Spectrum Strikes Back: Infrared POV Attacks on Traffic Sign Classification
AI Security, Red Teaming & Defense 2026
2026-06-29
[ "cs.CR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Traffic sign classification is a crucial task for autonomous vehicles, and numerous attacks against it have been identified. A majority of physical adversarial attacks involve attaching patches to traffic signs or projecting perturbations on them. Wh
Michael Kühr
5
[ "Michael Kühr", "Mevlüt Yildirim", "Maximilian Luedecke", "Mohammad Hamad", "Sebastian Steinhorst" ]
[]
http://arxiv.org/abs/2606.30153v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/pytorch/vision" ]
17,860
7,248
2026-08-13
BSD-3-Clause
154.05
Traffic sign classification is a crucial task for autonomous vehicles, and numerous attacks against it have been identified. A majority of physical adversarial attacks involve attaching patches to traffic signs or projecting perturbations on them. While they demonstrate high effectiveness, they are perceptible to human...
[ 0.012880000285804272, -0.029500000178813934, 0.026599999517202377, 0.040031999349594116, 0.03950599953532219, -0.0379519984126091, 0.0914589986205101, -0.10466399788856506, 0.04668800160288811, -0.02090499922633171, -0.007949000224471092, 0.0601079985499382, -0.013988000340759754, 0.004608...
[ -0.007218000013381243, -0.014692000113427639, 0.04113699868321419, 0.04707999899983406, 0.057861000299453735, -0.023462999612092972, 0.05319099873304367, -0.09538599848747253, -0.0075969998724758625, -0.066102996468544, -0.028996000066399574, -0.01600700058043003, 0.004141000099480152, -0....
AI Security, Red Teaming & Defense 2026
No code repository attached.
Traffic sign classification is a crucial task for autonomous vehicles, and numerous attacks against it have been identified.
We propose a persistence-of-vision-based attack that operates in the near-infrared light spectrum.
Our evaluation shows high attack success rates across our test scenarios.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:22:51.248639
2604.23067v1
Training a General Purpose Automated Red Teaming Model
AI Security, Red Teaming & Defense 2026
2026-04-24
[ "cs.CR", "cs.CL" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Automated methods for red teaming LLMs are an important tool to identify LLM vulnerabilities that may not be covered in static benchmarks, allowing for more thorough probing. They can also adapt to each specific LLM to discover weaknesses unique to i
Aishwarya Padmakumar
4
[ "Aishwarya Padmakumar", "Leon Derczynski", "Traian Rebedea", "Christopher Parisien" ]
[]
http://arxiv.org/abs/2604.23067v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/volcengine/verl", "https://github.com/NVIDIA/garak", "https://github.com/NVIDIA-NeMo/Skills" ]
22,947
0
Unknown
Unspecified
153.47
Automated methods for red teaming LLMs are an important tool to identify LLM vulnerabilities that may not be covered in static benchmarks, allowing for more thorough probing. They can also adapt to each specific LLM to discover weaknesses unique to it. Most current automated red teaming methods are intended for tacklin...
[ -0.002962999977171421, -0.04713800176978111, -0.051913999021053314, 0.0639289990067482, 0.005526999942958355, 0.03322000056505203, 0.04573500156402588, 0.02562900073826313, -0.06964600086212158, -0.059122998267412186, -0.047148000448942184, -0.0064860000275075436, 0.06062600016593933, 0.00...
[ -0.021227000281214714, -0.018377000465989113, -0.04877899959683418, 0.024140000343322754, 0.06475900113582611, 0.024981999769806862, 0.0006680000224150717, 0.0019509999547153711, -0.09743700176477432, -0.029634999111294746, -0.034217000007629395, -0.014196000061929226, 0.084818996489048, -...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Automated methods for red teaming LLMs are an important tool to identify LLM vulnerabilities that may not be covered in static benchmarks, allowing for more thorough probing.
Most current automated red teaming methods are intended for tackling safety and content moderation.
We demonstrate that finetuning small models, such as Qwen3-8B, using this pipeline results in a substantial improvement in their ability to generate attacks for both in and out of domain adversarial goals.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:45:21.301342
2605.27164v1
Query Symbolically or Retrieve Semantically? A Dataset and Method for Semi-Structured Question Answering
Enterprise RAG & Vector Search 2026
2026-05-26
[ "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Retrieval-Augmented Generation (RAG) systems for question answering typically retrieve evidence by semantic similarity between the query and document chunks. While effective for unstructured text, this approach is less reliable on semi-structured cor
Mateusz Czyżnikiewicz
9
[ "Mateusz Czyżnikiewicz", "Ryszard Tuora", "Adam Kozakiewicz", "Tomasz Ziętkiewicz", "Mateusz Galiński", "Michał Godziszewski", "Michał Karpowicz", "Timothy Hospedales", "Cristina Cornelio" ]
[]
http://arxiv.org/abs/2605.27164v1
VERIFIED_LIVE
https://github.com/pydantic/pydantic-ai
[ "https://github.com/yxh-y/TableRAG", "https://github.com/pydantic/pydantic-ai", "https://github.com/corneliocristina/DualGraphRAG", "https://github.com/jiaruzouu/T-RAG" ]
19,264
0
Unknown
Unspecified
153.17
Retrieval-Augmented Generation (RAG) systems for question answering typically retrieve evidence by semantic similarity between the query and document chunks. While effective for unstructured text, this approach is less reliable on semi-structured corpora where answering may require exact filtering, aggregation, or exha...
[ 0.03043700009584427, 0.049995001405477524, -0.027009999379515648, 0.015081999823451042, -0.04949900135397911, 0.029032999649643898, 0.09278199821710587, 0.042514998465776443, -0.0016380000161007047, -0.002675000112503767, 0.039351001381874084, -0.0719669982790947, 0.04804600030183792, 0.00...
[ -0.04664300009608269, 0.04661700129508972, 0.027475999668240547, 0.02407900057733059, 0.03057599999010563, 0.03275199979543686, 0.018164999783039093, 0.038995999842882156, -0.014503000304102898, -0.008186000399291515, -0.004137999843806028, -0.057117000222206116, 0.044773001223802567, -0.0...
Enterprise RAG & Vector Search 2026
git clone https://github.com/pydantic/pydantic-ai
Retrieval-Augmented Generation (RAG) systems for question answering typically retrieve evidence by semantic similarity between the query and document chunks.
We address this gap with DualGraph, a RAG framework that represents documents through two complementary views: a Textual Knowledge Graph for semantic retrieval and a Symbolic Knowledge Graph for symbolic querying over typed subject--predicate--object triples.
Experiments show that DualGraph consistently outperforms state-of-the-art dense-retrieval, GraphRAG, symbolic, and table-oriented baselines across question types.Code and data are available at https://github.com/corneliocristina/DualGraphRAG.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:40.607508
2606.13801v1
Neural Variability Enhances Artificial Network Robustness
AI Security, Red Teaming & Defense 2026
2026-06-11
[ "cs.LG", "q-bio.NC" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Neural responses in cortex exhibit substantial trial-to-trial variability in response to repeated stimuli, while peripheral sensory neurons respond far more consistently, leading many to wonder whether stochasticity may carry meaning. Existing work h
Robin Preble
4
[ "Robin Preble", "Praveen Venkatesh", "Stefan Mihalas", "Kameron Decker Harris" ]
[]
http://arxiv.org/abs/2606.13801v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/Trusted-AI/adversarial-robustness-toolbox", "https://github.com/pytorch/vision", "https://github.com/glomerulus-lab/structured-noise-robustness" ]
17,860
7,248
2026-08-13
BSD-3-Clause
153.15
Neural responses in cortex exhibit substantial trial-to-trial variability in response to repeated stimuli, while peripheral sensory neurons respond far more consistently, leading many to wonder whether stochasticity may carry meaning. Existing work has argued that noise and signal correlations may be optimized for disc...
[ -0.0031469999812543392, -0.057697001844644547, 0.05655800178647041, 0.06968200206756592, -0.024104999378323555, 0.010142999701201916, -0.031099999323487282, 0.01712699979543686, -0.036632999777793884, -0.14016899466514587, -0.07446099817752838, 0.01691499911248684, 0.06931599974632263, -0....
[ -0.04621899873018265, -0.059053998440504074, 0.11050599813461304, 0.0641700029373169, 0.03677000105381012, 0.007966999895870686, 0.016986999660730362, -0.12818799912929535, -0.0018759999657049775, -0.029176000505685806, -0.060478001832962036, -0.03290000185370445, 0.024288000538945198, 0.0...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Neural responses in cortex exhibit substantial trial-to-trial variability in response to repeated stimuli, while peripheral sensory neurons respond far more consistently, leading many to wonder whether stochasticity may carry meaning.
Existing work has argued that noise and signal correlations may be optimized for discrimination in animals, whereas artificial neural network (ANN) studies have shown similar benefits of noise in machine learning tasks, although most ANN work has neglected the effects of correlations.
These results suggest that structured noise in ANN activations generally improves robustness, establishing a biologically plausible strategy for creating robust artificial neural networks that only relies on local information.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:26:36.649389
2605.26646v1
UnityMAS-O: A General RL Optimization Framework for LLM-Based Multi-Agent Systems
Enterprise RAG & Vector Search 2026
2026-05-26
[ "cs.AI", "cs.CL", "cs.MA" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
LLM-based multi-agent systems decompose complex tasks into interacting roles, but most remain manually orchestrated by prompts, tools, and control rules, while agents are rarely optimized through a unified reinforcement learning interface. Existing R
Yiqun Chen
17
[ "Yiqun Chen", "Wei Yang", "Erhan Zhang", "Shijie Wang", "Qi Liu", "Zechun Niu", "Bin Zhang", "Haitao Li", "Rui Li", "Lingyong Yan", "Jinyuan Feng", "Biqing Qi", "Xiaochi Wei", "Yan Gao", "Yi Wu", "Yao Hu", "Jiaxin Mao" ]
[]
http://arxiv.org/abs/2605.26646v1
VERIFIED_LIVE
https://github.com/huggingface/trl
[ "https://github.com/chenyiqun/UnityMAS-O", "https://github.com/huggingface/trl" ]
19,066
0
Unknown
Unspecified
153.06
LLM-based multi-agent systems decompose complex tasks into interacting roles, but most remain manually orchestrated by prompts, tools, and control rules, while agents are rarely optimized through a unified reinforcement learning interface. Existing RL post-training frameworks mainly target single-policy optimization an...
[ 0.025544000789523125, -0.009852999821305275, -0.04132299870252609, -0.051162999123334885, 0.008112999610602856, 0.010528000071644783, -0.022564999759197235, 0.003135999897494912, -0.014365999959409237, 0.021606000140309334, -0.06991299986839294, 0.007083999924361706, 0.048840999603271484, ...
[ -0.01475400011986494, -0.03458999842405319, -0.08207499980926514, -0.012054000049829483, -0.019015999510884285, 0.012005000375211239, -0.02409599907696247, -0.02130899950861931, -0.04185299947857857, -0.007852000184357166, -0.0943870022892952, -0.048448000103235245, 0.05518899857997894, 0....
Enterprise RAG & Vector Search 2026
git clone https://github.com/huggingface/trl
LLM-based multi-agent systems decompose complex tasks into interacting roles, but most remain manually orchestrated by prompts, tools, and control rules, while agents are rarely optimized through a unified reinforcement learning interface.
Existing RL post-training frameworks mainly target single-policy optimization and lack abstractions for user-defined multi-agent workflows, structured interaction, role-specific credit assignment, and configurable parameter sharing.
These results show that UnityMAS-O can serve as a reusable substrate for converting diverse LLM-based multi-agent workflows into trainable multi-agent RL systems.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:36.908127
2607.04391v1
Memory-Orchestrated Semantic System (MOSS): An Auditable Agentic Memory Architecture
Enterprise RAG & Vector Search 2026
2026-07-05
[ "cs.CL" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Long-term memory remains a structural weakness of AI agents. The dominant approach, retrieval-augmented generation (RAG), relies on embedding-based similarity search, which is opaque by construction, difficult to audit, and bounded by the theoretical
Serge Lacasse
3
[ "Serge Lacasse", "Jérémie Hatier", "Alex Baker" ]
[]
http://arxiv.org/abs/2607.04391v1
VERIFIED_LIVE
https://github.com/MemoriLabs/Memori
[ "https://github.com/MemoriLabs/Memori" ]
15,782
2,955
2026-08-12
NOASSERTION
153
Long-term memory remains a structural weakness of AI agents. The dominant approach, retrieval-augmented generation (RAG), relies on embedding-based similarity search, which is opaque by construction, difficult to audit, and bounded by the theoretical limits of vector representations. We present the Memory-Orchestrated ...
[ 0.03320400044322014, 0.01514699961990118, -0.10960199683904648, 0.014643999747931957, -0.0023849999997764826, -0.0018749999580904841, 0.04729500040411949, -0.0742650032043457, 0.07171600311994553, 0.025676999241113663, 0.012575999833643436, 0.007197999861091375, 0.1137159988284111, 0.04553...
[ 0.0070039997808635235, -0.07360800355672836, -0.08294700086116791, 0.053523000329732895, -0.00977300014346838, 0.015714000910520554, -0.04063799977302551, -0.04093100130558014, 0.030174000188708305, -0.02059200033545494, -0.021539999172091484, 0.03568499907851219, 0.139491006731987, 0.0031...
Enterprise RAG & Vector Search 2026
git clone https://github.com/MemoriLabs/Memori
Long-term memory remains a structural weakness of AI agents.
The dominant approach, retrieval-augmented generation (RAG), relies on embedding-based similarity search, which is opaque by construction, difficult to audit, and bounded by the theoretical limits of vector representations.
We report on a longitudinal deployment unique in the agentic-memory literature: a year of continuous production over an individual scholar's working corpus--a conversational corpus reaching back to October 2024 (some 44 million tokens, retroactively indexed) comprising 110,183 segments, alongside 163,494 catalogued doc...
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:00.989579
2605.12565v1
Persona-Conditioned Adversarial Prompting (PCAP): Multi-Identity Red-Teaming for Enhanced Adversarial Prompt Discovery
AI Security, Red Teaming & Defense 2026
2026-05-12
[ "cs.CR" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Existing automated red-teaming pipelines often miss attacks that depend on attacker identity, framing, or multi-turn tactics. This under-coverage underestimates real-world risk. We introduce Persona-Conditioned Adversarial Prompting (PCAP), which con
Cristian Morasso
4
[ "Cristian Morasso", "Anisa Halimi", "Muhammad Zaid Hameed", "Douglas Leith" ]
[]
http://arxiv.org/abs/2605.12565v1
VERIFIED_LIVE
Not Applicable
[ "https://github.com/joke2k/faker" ]
19,370
0
Unknown
Unspecified
152.53
Existing automated red-teaming pipelines often miss attacks that depend on attacker identity, framing, or multi-turn tactics. This under-coverage underestimates real-world risk. We introduce Persona-Conditioned Adversarial Prompting (PCAP), which conditions adversarial search on attacker personas and strategy cards and...
[ -0.07017900049686432, 0.013148000463843346, 0.005458000116050243, 0.019101999700069427, 0.08682099729776382, 0.025134999305009842, 0.07734999805688858, -0.04839000105857849, 0.02207699976861477, -0.07150500267744064, -0.007530999835580587, -0.12063200026750565, 0.051208000630140305, -0.044...
[ -0.05205700173974037, 0.024507999420166016, -0.03281699866056442, 0.029496999457478523, 0.06140099838376045, 0.025498999282717705, 0.02201800048351288, -0.015178999863564968, -0.007877999916672707, -0.03780899941921234, -0.033886998891830444, -0.03637300059199333, 0.04712099954485893, -0.0...
AI Security, Red Teaming & Defense 2026
No code repository attached.
Existing automated red-teaming pipelines often miss attacks that depend on attacker identity, framing, or multi-turn tactics.
We introduce Persona-Conditioned Adversarial Prompting (PCAP), which conditions adversarial search on attacker personas and strategy cards and runs parallel persona-conditioned beam searches to discover diverse, transferable jailbreaks.
PCAP is orthogonal to the underlying search algorithm and substantially increases attack success rate (ASR) and prompt diversity (e.g., ASR on GPT-OSS~120B from $\approx58\% \rightarrow \approx97\%$), improving attack strategy coverage and diversity.
1
AI Security, Red Teaming & Defense 2026 Research
Explosive (>50/mo)
208
2026-08-13T17:38:31.567390
2606.02643v1
Inference Cost Attacks for Retrieval-Augmented Large Language Models
Enterprise RAG & Vector Search 2026
2026-05-31
[ "cs.CR", "cs.AI", "cs.DB" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Retrieval-Augmented Generation (RAG)-enhanced LLM systems, while powerful, introduce substantial inference costs due to the inclusion of an extra multi-stage pipeline that dynamically retrieves and synthesizes information from external knowledge sour
Chengliang Liu
4
[ "Chengliang Liu", "Liangbo Ning", "Yujuan Ding", "Wenqi Fan" ]
[]
http://arxiv.org/abs/2606.02643v1
VERIFIED_LIVE
https://github.com/GoogleCloudPlatform/generative-ai
[ "https://github.com/GoogleCloudPlatform/generative-ai" ]
17,584
0
Unknown
Unspecified
152.43
Retrieval-Augmented Generation (RAG)-enhanced LLM systems, while powerful, introduce substantial inference costs due to the inclusion of an extra multi-stage pipeline that dynamically retrieves and synthesizes information from external knowledge sources. This high operational cost exposes a critical vulnerability to In...
[ -0.10615099966526031, -0.049254998564720154, 0.032196998596191406, 0.06675200164318085, 0.07238899916410446, 0.10892900079488754, 0.077674001455307, -0.03135799989104271, 0.10153499990701675, 0.006647000089287758, 0.03614199906587601, -0.024816999211907387, 0.07120899856090546, -0.02618199...
[ -0.11558499932289124, -0.029178999364376068, -0.013261999934911728, 0.056070998311042786, 0.07206299901008606, 0.008391999639570713, 0.01958799920976162, -0.009526999667286873, 0.06169699877500534, 0.025428999215364456, -0.018245000392198563, -0.017676999792456627, 0.1392049938440323, -0.0...
Enterprise RAG & Vector Search 2026
git clone https://github.com/GoogleCloudPlatform/generative-ai
Retrieval-Augmented Generation (RAG)-enhanced LLM systems, while powerful, introduce substantial inference costs due to the inclusion of an extra multi-stage pipeline that dynamically retrieves and synthesizes information from external knowledge sources.
We argue that a more feasible and potent threat to RAG-enhanced LLM systems arises from poisoning external knowledge bases (e.g., web knowledge from the Internet).
Extensive experiments across three real-world datasets demonstrate that RA-ICA increases token consumption by up to 13.12 times with an over 90% success rate, without degrading the integrity of the generated answer.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:29.522439
2605.27445v1
RAGe: A Retrieval-Augmented Generation Evaluation Framework
Enterprise RAG & Vector Search 2026
2026-05-23
[ "cs.IR", "cs.AI" ]
ArXiv Standard
http://arxiv.org/licenses/nonexclusive-distrib/1.0/
true
85
Enterprise Safe
ArXiv Open Access
0
0
1
Deploying Large Language Model (LLM) applications, particularly those relying on Retrieval-Augmented Generation (RAG), remains challenging due to high computational demands, outdated knowledge bases, and the need to manually select optimal pipeline c
Larissa Guder
8
[ "Larissa Guder", "João Pedro de Moura", "Arthur Accorsi", "Gustavo Losch do Amaral", "Maurício Cecílio Magnaguagno", "Felipe Meneguzzi", "Marcio Sorraglia Pinho", "Dalvan Griebler" ]
[]
http://arxiv.org/abs/2605.27445v1
VERIFIED_LIVE
https://github.com/confident-ai/deepeval
[ "https://github.com/confident-ai/deepeval", "https://github.com/DocAILab/XRAG", "https://github.com/giampaolo/psutil", "https://github.com/castorini/ragnarok" ]
17,570
0
Unknown
Unspecified
152.02
Deploying Large Language Model (LLM) applications, particularly those relying on Retrieval-Augmented Generation (RAG), remains challenging due to high computational demands, outdated knowledge bases, and the need to manually select optimal pipeline components. In this work, we propose a modular framework for benchmarki...
[ -0.06261000037193298, 0.0053090001456439495, -0.07398699969053268, 0.050870999693870544, 0.013981999829411507, 0.09802500158548355, 0.018431000411510468, -0.016946999356150627, 0.02598699927330017, -0.040130000561475754, 0.04128799960017204, -0.08070799708366394, 0.06666900217533112, -0.03...
[ -0.04593700170516968, -0.04901599884033203, -0.0027910000644624233, 0.013605999760329723, 0.0019460000330582261, -0.011327000334858894, -0.056745000183582306, 0.013907000422477722, -0.03073900006711483, -0.06544899940490723, -0.016958000138401985, -0.0045900000259280205, 0.05398799851536751,...
Enterprise RAG & Vector Search 2026
git clone https://github.com/confident-ai/deepeval
Deploying Large Language Model (LLM) applications, particularly those relying on Retrieval-Augmented Generation (RAG), remains challenging due to high computational demands, outdated knowledge bases, and the need to manually select optimal pipeline components.
In this work, we propose a modular framework for benchmarking and guiding the efficient development of RAG applications by focusing on resource telemetry and component recommendation, suggesting the best components for a domain-specific dataset.
Our approach leverages core techniques in LLM applications, including document chunking, vector databases, embedding models, and retrievers, to evaluate trade-offs among accuracy, efficiency, and scalability.
1
Enterprise RAG & Vector Search 2026 Research
Explosive (>50/mo)
232
2026-08-13T18:23:47.926286