row_id
string
section
string
section_slug
string
resource_type
string
marker
string
title
string
url
string
url_kind
string
domain
string
annotation
string
description
string
key_contribution
string
novelty
string
impact
string
signal
string
signal_strength
string
source_readme
string
source_line
int64
source_url
string
date_added
string
collection
string
collection_slug
string
user_goal
string
lifecycle_stages
string
audience
string
loop_layer
string
scope_fit
string
evidence_class
string
evidence_tier
string
source_status
string
canonical_url
string
source_title
string
source_description
string
authors
string
publication_date
string
publication_year
string
publication_venue
string
publisher
string
doi
string
publication_note
string
primary_category
string
metadata_source
string
github_repo
string
github_stars
string
github_forks
string
github_license
string
github_created_at
string
github_updated_at
string
arxiv_id
string
audited_at
timestamp[ms]
ale-0401
Verification And Feedback Gates
verification-and-feedback-gates
Paper
📄
Self-Authored Verification Is Unreliable in Heuristic Self-Improving Agents
https://arxiv.org/abs/2607.24300
external
arxiv.org
Names and measures the verifier-deployment gap: when a self-improving agent controls both the artifact it rewrites and the tests that judge the rewrite, self-assigned scores stay near-perfect while sealed deployment performance degrades. Introduces a Sealed Exogenous evaluation protocol and asks the practical question ...
Names and measures the verifier-deployment gap: when a self-improving agent controls both the artifact it rewrites and the tests that judge the rewrite, self-assigned scores stay near-perfect while sealed deployment performance degrades. Introduces a Sealed Exogenous evaluation protocol and asks the practical question ...
Names and measures the verifier-deployment gap: when a self-improving agent controls both the artifact it rewrites and the tests that judge the rewrite, self-assigned scores stay near-perfect while sealed deployment performance degrades. Introduces a Sealed Exogenous evaluation protocol and asks the practical question ...
Verification is promoted from a final check to a loop-control signal. Names and measures the verifier-deployment gap: when a self-improving agent controls both the artifact it rewrites and the tests that judge the rewrite, self-assigned scores stay near-perfect while sealed deployment performance degrades. Introduces a...
Use Self-Authored Verification Is Unreliable in Heuristic Self-Improving Agents to measure progress and gate completion with repeatable evidence.
Research source arXiv:2607.24300; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,049
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1049
2026-07-28
Verify
verify
Gate progress with tests, evals, and evidence.
verification;exit
researcher;evaluator
harness
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.24300
[2607.24300] Self-Authored Verification Is Unreliable in Heuristic Self-Improving Agents
Self-improving agents accumulate capability by repeatedly rewriting procedural policies, controllers, or heuristic rules. They typically rely on self-authored tests or metrics to decide whether to accept subsequent edits. The agent controls both the optimized object and its verifier. As a result, self-assigned scores c...
Diandian Guo; Cong Cao; Fangfang Yuan; Yingqi Wang; Yueshan Wang; Dakui Wang
2026-07-27
2026
arXiv
arXiv
9 pages, 6 figures
cs.CL
arxiv-api
2607.24300
2026-07-29T08:06:54
ale-0402
Verification And Feedback Gates
verification-and-feedback-gates
Paper
📄
Falsifiable Commitment Planning for Self-Correcting Web Agents
https://arxiv.org/abs/2607.24167
external
arxiv.org
Long-horizon web agents drift: a trajectory stays locally plausible long after the state or plan assumption stopped supporting the instruction. FCPAgent makes each plan step a Falsifiable Commitment Unit -- a subgoal grounded in a reusable skill plus confirming evidence, falsifying evidence, and a confidence score -- a...
Long-horizon web agents drift: a trajectory stays locally plausible long after the state or plan assumption stopped supporting the instruction. FCPAgent makes each plan step a Falsifiable Commitment Unit -- a subgoal grounded in a reusable skill plus confirming evidence, falsifying evidence, and a confidence score -- a...
Long-horizon web agents drift: a trajectory stays locally plausible long after the state or plan assumption stopped supporting the instruction. FCPAgent makes each plan step a Falsifiable Commitment Unit -- a subgoal grounded in a reusable skill plus confirming evidence, falsifying evidence, and a confidence score -- a...
The work targets tasks that exceed a single context window or prompt session. Long-horizon web agents drift: a trajectory stays locally plausible long after the state or plan assumption stopped supporting the instruction. FCPAgent makes each plan step a Falsifiable Commitment Unit -- a subgoal grounded in a reusable sk...
Use Falsifiable Commitment Planning for Self-Correcting Web Agents to measure progress and gate completion with repeatable evidence.
Research source arXiv:2607.24167; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,050
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1050
2026-07-28
Verify
verify
Gate progress with tests, evals, and evidence.
verification;state;exit
researcher;evaluator
harness
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.24167
[2607.24167] Falsifiable Commitment Planning for Self-Correcting Web Agents
Long-horizon web agents often go off track before final failure: a trajectory can remain locally plausible even after the current state, reused skill, or plan assumption no longer supports the user instruction. Existing agents can plan, reflect, or reuse experience, but their plans rarely specify the evidence under whi...
Guangyi Liu; Huan Zhao; Quanming Yao
2026-07-27
2026
arXiv
arXiv
cs.AI
arxiv-api
2607.24167
2026-07-29T08:06:54
ale-0403
Verification And Feedback Gates
verification-and-feedback-gates
Paper
📄
Adversarial Test-Hardening for AI-Written Code: An Instrument Autopsy and a Pre-Registered Causal Estimate of the Critic Loop
https://arxiv.org/abs/2607.23002
external
arxiv.org
A Tester model writes tests, mutation testing names surviving injected defects, and a Critic model writes tests to kill exactly those -- every verdict decided mechanically so no model judges another. The loop killed 105 mutants one-shot generation missed and lost none. The more valuable contribution is the autopsy: an ...
A Tester model writes tests, mutation testing names surviving injected defects, and a Critic model writes tests to kill exactly those -- every verdict decided mechanically so no model judges another. The loop killed 105 mutants one-shot generation missed and lost none. The more valuable contribution is the autopsy: an ...
A Tester model writes tests, mutation testing names surviving injected defects, and a Critic model writes tests to kill exactly those -- every verdict decided mechanically so no model judges another. The loop killed 105 mutants one-shot generation missed and lost none. The more valuable contribution is the autopsy: an ...
Evaluation data is used as the feedback signal for improving loop behavior. A Tester model writes tests, mutation testing names surviving injected defects, and a Critic model writes tests to kill exactly those -- every verdict decided mechanically so no model judges another. The loop killed 105 mutants one-shot generat...
Use Adversarial Test-Hardening for AI-Written Code: An Instrument Autopsy and a Pre-Registered Causal Estimate of the Critic Loop to measure progress and gate completion with repeatable evidence.
Research source arXiv:2607.23002; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,051
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1051
2026-07-28
Verify
verify
Gate progress with tests, evals, and evidence.
verification
researcher;evaluator
harness
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.23002
[2607.23002] Adversarial Test-Hardening for AI-Written Code: An Instrument Autopsy and a Pre-Registered Causal Estimate of the Critic Loop
Large language models increasingly write both code and the tests meant to check it; coverage records what ran, not what was verified. We study an adversarial test-hardening loop under a mechanical oracle: a Tester model writes tests, mutation testing names surviving injected defects, and a Critic model writes tests to ...
Jeff Otterson
2026-07-25
2026
arXiv
arXiv
26 pages. Two pre-registered experiments; protocols, all run receipts, and analysis code at https://github.com/Jott2121/crucible
cs.SE
arxiv-api
2607.23002
2026-07-29T08:06:54
ale-0404
Verification And Feedback Gates
verification-and-feedback-gates
Paper
📄
Beyond Aggregate Risk: Role-Stratified Conformal Risk Control for LLM Tool Calls
https://arxiv.org/abs/2607.24343
external
arxiv.org
Tool-call arguments carry wildly different risk -- untrusted content may safely shape an email body but must never determine a recipient, account, command, or credential -- yet standard statistical control certifies the action as a whole, letting rare high-risk field failures hide behind benign arguments. Adds a calibr...
Tool-call arguments carry wildly different risk -- untrusted content may safely shape an email body but must never determine a recipient, account, command, or credential -- yet standard statistical control certifies the action as a whole, letting rare high-risk field failures hide behind benign arguments. Adds a calibr...
Tool-call arguments carry wildly different risk -- untrusted content may safely shape an email body but must never determine a recipient, account, command, or credential -- yet standard statistical control certifies the action as a whole, letting rare high-risk field failures hide behind benign arguments. Adds a calibr...
Untrusted intake is treated as a loop-level security boundary. Tool-call arguments carry wildly different risk -- untrusted content may safely shape an email body but must never determine a recipient, account, command, or credential -- yet standard statistical control certifies the action as a whole, letting rare high-...
Use Beyond Aggregate Risk: Role-Stratified Conformal Risk Control for LLM Tool Calls to measure progress and gate completion with repeatable evidence.
Research source arXiv:2607.24343; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,052
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1052
2026-07-28
Verify
verify
Gate progress with tests, evals, and evidence.
workspace;budget
researcher;evaluator
harness
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.24343
[2607.24343] Beyond Aggregate Risk: Role-Stratified Conformal Risk Control for LLM Tool Calls
Language-model agents act through structured tool calls whose arguments carry different risks. Untrusted content may safely influence an email body but should not determine a recipient, account, command, or credential. Existing statistical methods typically control risk over the entire action, allowing failures in rare...
Md Ashikur Rahman; Md Arifur Rahman; Niamul Hassan Samin; Khandaker Rifah Tasnia; Sifat Rahman Ahona; Juena Ahmed Noshin
2026-07-27
2026
arXiv
arXiv
cs.LG
arxiv-api
2607.24343
2026-07-29T08:06:54
ale-0405
Verification And Feedback Gates
verification-and-feedback-gates
Paper
📄
From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards for Open-Ended LLM Self-Improvement
https://arxiv.org/abs/2607.23802
external
arxiv.org
RLVR works where correctness is deterministically checkable and stalls everywhere else, leaving open-ended tasks dependent on preference data, reward models, or LLM judges with their bias, capability ceilings, and inference cost. RLSVR borrows the self-supervised trick of constructing pretext tasks: transform an open-e...
RLVR works where correctness is deterministically checkable and stalls everywhere else, leaving open-ended tasks dependent on preference data, reward models, or LLM judges with their bias, capability ceilings, and inference cost. RLSVR borrows the self-supervised trick of constructing pretext tasks: transform an open-e...
RLVR works where correctness is deterministically checkable and stalls everywhere else, leaving open-ended tasks dependent on preference data, reward models, or LLM judges with their bias, capability ceilings, and inference cost. RLSVR borrows the self-supervised trick of constructing pretext tasks: transform an open-e...
Verification is promoted from a final check to a loop-control signal. RLVR works where correctness is deterministically checkable and stalls everywhere else, leaving open-ended tasks dependent on preference data, reward models, or LLM judges with their bias, capability ceilings, and inference cost. RLSVR borrows the se...
Use From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards for Open-Ended LLM Self-Improvement to measure progress and gate completion with repeatable evidence.
Research source arXiv:2607.23802; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,053
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1053
2026-07-28
Verify
verify
Gate progress with tests, evals, and evidence.
verification;budget
researcher;evaluator
harness
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.23802
[2607.23802] From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards for Open-Ended LLM Self-Improvement
Reinforcement Learning with Verifiable Rewards (RLVR) has driven recent progress in reasoning-oriented large language models (LLMs) by enabling large-scale optimization. However, its applicability remains largely limited to domains such as mathematics and coding, where correctness can be deterministically verified. Ope...
Qinsi Wang; Jing Shi; Huazheng Wang; Kun Wan; Yiran Wu; Bo Liu; Qingyun Wu; Hai Helen Li; Yiran Chen; Handong Zhao; Wentian Zhao
2026-07-26
2026
arXiv
arXiv
COLM 2026
cs.AI
arxiv-api
2607.23802
2026-07-29T08:06:54
ale-0406
Verification And Feedback Gates
verification-and-feedback-gates
Paper
📄
A Frozen 12B Beats Frontier Models on Verified Work: 100% Accuracy, 0 Tokens, Bit-Exact, Forever
https://arxiv.org/abs/2607.23806
external
arxiv.org
Verification-gated memoization taken to its limit: the model stays frozen while a persistent store accumulates only those solutions that passed an independent verification step, so later encounters with a seen task replay a checked answer instead of re-deriving it. Read the headline accuracy and token figures as proper...
Verification-gated memoization taken to its limit: the model stays frozen while a persistent store accumulates only those solutions that passed an independent verification step, so later encounters with a seen task replay a checked answer instead of re-deriving it. Read the headline accuracy and token figures as proper...
Verification-gated memoization taken to its limit: the model stays frozen while a persistent store accumulates only those solutions that passed an independent verification step, so later encounters with a seen task replay a checked answer instead of re-deriving it. Read the headline accuracy and token figures as proper...
Durable execution and replay are treated as first-class loop infrastructure. Verification-gated memoization taken to its limit: the model stays frozen while a persistent store accumulates only those solutions that passed an independent verification step, so later encounters with a seen task replay a checked answer inst...
Use A Frozen 12B Beats Frontier Models on Verified Work: 100% Accuracy, 0 Tokens, Bit-Exact, Forever to measure progress and gate completion with repeatable evidence.
Research source arXiv:2607.23806; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,054
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1054
2026-07-28
Verify
verify
Gate progress with tests, evals, and evidence.
verification;state;budget
researcher;evaluator
harness
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.23806
[2607.23806] A Frozen 12B Beats Frontier Models on Verified Work: 100% Accuracy, 0 Tokens, Bit-Exact, Forever
Improving a language model today means retraining it: enormous compute, a new opaque model each cycle, non-deterministic output. We take the opposite path: the model stays frozen, and a persistent memory of verified solutions grows beside it. Once a problem family is solved and has passed an independent verification st...
Sietse Schelpe
2026-07-26
2026
arXiv
arXiv
Industry experience report. 14 pages, 8 figures. Public testbench: https://corbenic-galahad-bench.hf.space; companion repository with SHA-256 provenance manifest: https://github.com/corbenicai/galahad-bench
cs.CL
arxiv-api
2607.23806
2026-07-29T08:06:54
ale-0407
Securing Unattended Loops
securing-unattended-loops
Critique
⚠️
The lethal trifecta for AI agents
https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/
external
simonwillison.net
Simon Willison's rule of thumb: private data, untrusted content, and an exfiltration channel must never meet inside one unattended agent.
Simon Willison's rule of thumb: private data, untrusted content, and an exfiltration channel must never meet inside one unattended agent.
Simon Willison's rule of thumb: private data, untrusted content, and an exfiltration channel must never meet inside one unattended agent.
Untrusted intake is treated as a loop-level security boundary. Simon Willison's rule of thumb: private data, untrusted content, and an exfiltration channel must never meet inside one unattended agent.
Use The lethal trifecta for AI agents to bound risk before recurring or unattended execution.
Contextual source from simonwillison.net; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,062
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1062
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;budget;escalation
operator;security
operations
enabling
risk-analysis
B
ok
https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/
The lethal trifecta for AI agents: private data, untrusted content, and external communication
If you are a user of LLM systems that use tools (you can call them “AI agents” if you like) it is critically important that you understand the risk of …
Simon Willison
2025
Simon Willison’s Weblog
html-meta
2026-07-29T08:06:54
ale-0408
Securing Unattended Loops
securing-unattended-loops
Critique
⚠️
Prompt injection series
https://simonwillison.net/series/prompt-injection/
external
simonwillison.net
Ongoing series on the core unsolved vulnerability for loops whose intake includes content written by strangers.
Ongoing series on the core unsolved vulnerability for loops whose intake includes content written by strangers.
Ongoing series on the core unsolved vulnerability for loops whose intake includes content written by strangers.
Untrusted intake is treated as a loop-level security boundary. Ongoing series on the core unsolved vulnerability for loops whose intake includes content written by strangers.
Use Prompt injection series to bound risk before recurring or unattended execution.
Contextual source from simonwillison.net; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,063
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1063
Govern
govern
Bound permissions, cost, failure, and escalation.
intake
operator;security
operations
enabling
risk-analysis
B
ok
https://simonwillison.net/series/prompt-injection/
Simon Willison: Prompt injection
Simon Willison
Simon Willison’s Weblog
html-meta
2026-07-29T08:06:54
ale-0409
Securing Unattended Loops
securing-unattended-loops
Docs
📚
Agentic AI - Threats and Mitigations
https://genai.owasp.org/resource/agentic-ai-threats-and-mitigations/
external
genai.owasp.org
OWASP threat model for agentic systems, useful when reviewing intake, memory, tool, and delegation boundaries.
OWASP threat model for agentic systems, useful when reviewing intake, memory, tool, and delegation boundaries.
OWASP threat model for agentic systems, useful when reviewing intake, memory, tool, and delegation boundaries.
Persistent memory is treated as an external runtime artifact. OWASP threat model for agentic systems, useful when reviewing intake, memory, tool, and delegation boundaries.
Use Agentic AI - Threats and Mitigations to bound risk before recurring or unattended execution.
Primary documentation from a platform, SDK, standard, or framework; strong implementation signal.
high
README.md
1,064
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1064
Govern
govern
Bound permissions, cost, failure, and escalation.
intake;workspace;context;delegation
builder;operator;security
operations
enabling
technical-documentation
A
ok
https://genai.owasp.org/resource/agentic-ai-threats-and-mitigations/
Agentic AI - OWASP Lists Threats and Mitigations
Explore key threats and mitigation strategies for agentic AI, focusing on security measures to address vulnerabilities in AI applications and their potential risks.
OWASPGenAIProject Editor
OWASP Gen AI Security Project
html-meta
2026-07-29T08:06:54
ale-0410
Securing Unattended Loops
securing-unattended-loops
Docs
📚
Designing AI agents to resist prompt injection
https://openai.com/index/designing-agents-to-resist-prompt-injection/
external
openai.com
OpenAI's official defense-in-depth guidance: least privilege, sandboxed tools, output verification, and human confirmation for the high-impact actions an unattended loop might take.
OpenAI's official defense-in-depth guidance: least privilege, sandboxed tools, output verification, and human confirmation for the high-impact actions an unattended loop might take.
OpenAI's official defense-in-depth guidance: least privilege, sandboxed tools, output verification, and human confirmation for the high-impact actions an unattended loop might take.
Primary-source operational guidance rather than commentary. OpenAI's official defense-in-depth guidance: least privilege, sandboxed tools, output verification, and human confirmation for the high-impact actions an unattended loop might take.
Use Designing AI agents to resist prompt injection to bound risk before recurring or unattended execution.
Primary documentation from a platform, SDK, standard, or framework; strong implementation signal.
high
README.md
1,065
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1065
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;verification;escalation
builder;operator;security
operations
enabling
technical-documentation
A
ok
https://openai.com/index/designing-agents-to-resist-prompt-injection/
Designing AI agents to resist prompt injection | OpenAI
How ChatGPT defends against prompt injection and social engineering by constraining risky actions and protecting sensitive data in agent workflows.
OpenAI
html-meta
2026-07-29T08:06:54
ale-0411
Securing Unattended Loops
securing-unattended-loops
Tool
🧰
sandbox-runtime
https://github.com/anthropic-experimental/sandbox-runtime
external
github.com
Anthropic's OS-level filesystem and network sandboxing for arbitrary processes without requiring a container.
Anthropic's OS-level filesystem and network sandboxing for arbitrary processes without requiring a container.
Anthropic's OS-level filesystem and network sandboxing for arbitrary processes without requiring a container.
Execution isolation and permission boundaries are part of the design. Anthropic's OS-level filesystem and network sandboxing for arbitrary processes without requiring a container.
Use sandbox-runtime to bound risk before recurring or unattended execution.
Inspectable GitHub source (4,783 stars; 377 forks; Apache-2.0 license; updated 2026-07-28); popularity is context, not proof of reliability.
medium
README.md
1,066
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1066
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace
builder;operator;security
operations
enabling
source-implementation
A
ok
https://github.com/anthropic-experimental/sandbox-runtime
GitHub - anthropic-experimental/sandbox-runtime: A lightweight sandboxing tool for enforcing filesystem and network restrictions on arbitrary processes at the OS level, without requiring a container. · GitHub
A lightweight sandboxing tool for enforcing filesystem and network restrictions on arbitrary processes at the OS level, without requiring a container. - anthropic-experimental/sandbox-runtime
2025-10-20
2025
anthropic-experimental/sandbox-runtime
GitHub
github-api
anthropic-experimental/sandbox-runtime
4783
377
Apache-2.0
2025-10-20T02:52:10Z
2026-07-28T14:51:22Z
2026-07-29T08:06:54
ale-0412
Securing Unattended Loops
securing-unattended-loops
Tool
🧰
E2B
https://github.com/e2b-dev/E2B
external
github.com
Open-source isolated cloud sandboxes for running untrusted, AI-generated code inside agent loops.
Open-source isolated cloud sandboxes for running untrusted, AI-generated code inside agent loops.
Open-source isolated cloud sandboxes for running untrusted, AI-generated code inside agent loops.
Execution isolation and permission boundaries are part of the design. Open-source isolated cloud sandboxes for running untrusted, AI-generated code inside agent loops.
Use E2B to bound risk before recurring or unattended execution.
Inspectable GitHub source (13,164 stars; 977 forks; Apache-2.0 license; updated 2026-07-29); popularity is context, not proof of reliability.
medium
README.md
1,067
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1067
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace
builder;operator;security
operations
enabling
source-implementation
A
ok
https://github.com/e2b-dev/E2B
GitHub - e2b-dev/E2B: Open-source, secure environment with real-world tools for enterprise-grade agents. · GitHub
Open-source, secure environment with real-world tools for enterprise-grade agents. - e2b-dev/E2B
2023-03-04
2023
e2b-dev/E2B
GitHub
github-api
e2b-dev/E2B
13164
977
Apache-2.0
2023-03-04T13:41:18Z
2026-07-29T07:45:35Z
2026-07-29T08:06:54
ale-0413
Securing Unattended Loops
securing-unattended-loops
Docs
📚
Modal Sandboxes
https://modal.com/docs/guide/sandboxes
external
modal.com
Secure sandboxed execution for agent-driven code with resource limits and network controls.
Secure sandboxed execution for agent-driven code with resource limits and network controls.
Secure sandboxed execution for agent-driven code with resource limits and network controls.
Execution isolation and permission boundaries are part of the design. Secure sandboxed execution for agent-driven code with resource limits and network controls.
Use Modal Sandboxes to bound risk before recurring or unattended execution.
Primary documentation from a platform, SDK, standard, or framework; strong implementation signal.
high
README.md
1,068
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1068
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace
builder;operator;security
operations
enabling
technical-documentation
A
ok
https://modal.com/docs/guide/sandboxes
Sandboxes | Modal Docs
This page is a high-level guide to Sandboxes, secure containers for executing untrusted user or agent code on Modal.
Modal
html-meta
2026-07-29T08:06:54
ale-0414
Securing Unattended Loops
securing-unattended-loops
Tool
🧰
Daytona
https://www.daytona.io/
external
www.daytona.io
Infrastructure for running AI-generated code in fast, isolated sandboxes.
Infrastructure for running AI-generated code in fast, isolated sandboxes.
Infrastructure for running AI-generated code in fast, isolated sandboxes.
Execution isolation and permission boundaries are part of the design. Infrastructure for running AI-generated code in fast, isolated sandboxes.
Use Daytona to bound risk before recurring or unattended execution.
Working implementation, framework, runtime, or repository; signal comes from usable code and ecosystem adoption.
high
README.md
1,069
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1069
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace
builder;operator;security
operations
enabling
implementation
A
ok
https://www.daytona.io/
Daytona - Secure Infrastructure for Running AI-Generated Code
Deploy Al code with confidence using Daytona's lightning-fast infrastructure. 90ms environment creation, stateful operations, and enterprise-grade security.
daytona.io
domain-fallback
2026-07-29T08:06:54
ale-0415
Securing Unattended Loops
securing-unattended-loops
Tool
🧰
peerd
https://github.com/NotASithLord/peerd
external
github.com
Browser-extension harness that runs the agent loop entirely client-side with user-supplied keys, sandboxed compute, and per-environment actor agents that hold only their tools and no API keys, isolating the orchestrator from untrusted content as a prompt-injection boundary.
Browser-extension harness that runs the agent loop entirely client-side with user-supplied keys, sandboxed compute, and per-environment actor agents that hold only their tools and no API keys, isolating the orchestrator from untrusted content as a prompt-injection boundary.
Browser-extension harness that runs the agent loop entirely client-side with user-supplied keys, sandboxed compute, and per-environment actor agents that hold only their tools and no API keys, isolating the orchestrator from untrusted content as a prompt-injection boundary.
Orchestration and control flow are made explicit and inspectable. Browser-extension harness that runs the agent loop entirely client-side with user-supplied keys, sandboxed compute, and per-environment actor agents that hold only their tools and no API keys, isolating the orchestrator from untrusted content as a prompt...
Use peerd to bound risk before recurring or unattended execution.
Inspectable GitHub source (371 stars; 36 forks; Apache-2.0 license; updated 2026-07-28); popularity is context, not proof of reliability.
medium
README.md
1,070
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1070
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;delegation
builder;operator;security
operations
enabling
source-implementation
A
ok
https://github.com/NotASithLord/peerd
GitHub - NotASithLord/peerd: The first AI agent harness native to the browser. A browser extension that runs a full agent loop where you already work: it drives your tabs, spins up sandboxed compute (JS notebooks, WASM Linux VMs, client-side apps), and shares what it builds peer-to-peer. BYOK, no backend, no telemetry....
The first AI agent harness native to the browser. A browser extension that runs a full agent loop where you already work: it drives your tabs, spins up sandboxed compute (JS notebooks, WASM Linux VMs, client-side apps), and shares what it builds peer-to-peer. BYOK, no backend, no telemetry. - NotASithLord/peerd
2026-06-22
2026
NotASithLord/peerd
GitHub
github-api
NotASithLord/peerd
371
36
Apache-2.0
2026-06-22T05:55:13Z
2026-07-28T12:30:21Z
2026-07-29T08:06:54
ale-0416
Securing Unattended Loops
securing-unattended-loops
Paper
📄
When Claws Remember but Do Not Tell: Stealthy Memory Injection in Persistent Personal Agents
https://arxiv.org/abs/2607.05189
external
arxiv.org
Shows one poisoned email can write hidden entries into a persistent personal agent's long-term memory that silently alter future unattended runs, introducing the 108-case WhisperBench evaluation and the MemGhost attack that reaches 87.5% success.
Shows one poisoned email can write hidden entries into a persistent personal agent's long-term memory that silently alter future unattended runs, introducing the 108-case WhisperBench evaluation and the MemGhost attack that reaches 87.5% success.
Shows one poisoned email can write hidden entries into a persistent personal agent's long-term memory that silently alter future unattended runs, introducing the 108-case WhisperBench evaluation and the MemGhost attack that reaches 87.5% success.
Evaluation data is used as the feedback signal for improving loop behavior. Shows one poisoned email can write hidden entries into a persistent personal agent's long-term memory that silently alter future unattended runs, introducing the 108-case WhisperBench evaluation and the MemGhost attack that reaches 87.5% succes...
Use When Claws Remember but Do Not Tell: Stealthy Memory Injection in Persistent Personal Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.05189; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,071
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1071
Govern
govern
Bound permissions, cost, failure, and escalation.
context;verification;state
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.05189
[2607.05189] When Claws Remember but Do Not Tell: Stealthy Memory Injection in Persistent Personal Agents
Persistent personal agents combine long-term memory with access to users' external environments, enabling personalized foreground assistance and proactive background execution. This integration also creates a new path to compromise: untrusted external content can be silently written into persistent memory and later reu...
Yechao Zhang; Shiqian Zhao; Jiawen Zhang; Jie Zhang; Gelei Deng; Xiaogeng Liu; Chaowei Xiao; Tianwei Zhang
2026-07-06
2026
arXiv
arXiv
25 pages, 8 figures. Preprint
cs.CR
arxiv-api
2607.05189
2026-07-29T08:06:54
ale-0417
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Your Agent's Memories Are Not Its Own: Forged Reasoning Attacks on LLM Agent Memory and Defenses
https://arxiv.org/abs/2607.05029
external
arxiv.org
Introduces FARMA, an attack that plants forged reasoning traces in an agent's persistent memory so poisoned rationales carry into future runs, and SENTINEL, a reasoning-guard defense that cut attack success from up to 100% to zero in evaluation.
Introduces FARMA, an attack that plants forged reasoning traces in an agent's persistent memory so poisoned rationales carry into future runs, and SENTINEL, a reasoning-guard defense that cut attack success from up to 100% to zero in evaluation.
Introduces FARMA, an attack that plants forged reasoning traces in an agent's persistent memory so poisoned rationales carry into future runs, and SENTINEL, a reasoning-guard defense that cut attack success from up to 100% to zero in evaluation.
Evaluation data is used as the feedback signal for improving loop behavior. Introduces FARMA, an attack that plants forged reasoning traces in an agent's persistent memory so poisoned rationales carry into future runs, and SENTINEL, a reasoning-guard defense that cut attack success from up to 100% to zero in evaluation...
Use Your Agent's Memories Are Not Its Own: Forged Reasoning Attacks on LLM Agent Memory and Defenses to bound risk before recurring or unattended execution.
Research source arXiv:2607.05029; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,072
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1072
Govern
govern
Bound permissions, cost, failure, and escalation.
context;verification;state
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.05029
[2607.05029] Your Agent's Memories Are Not Its Own: Forged Reasoning Attacks on LLM Agent Memory and Defenses
Persistent memory has enabled large language model (LLM) agents to store factual knowledge, prior decisions, reasoning histories, tool usage information, and context. While this has improved the agent's functionality and continuity across tasks, it has also introduced a new attack surface: the agent's own reasoning his...
Neeraj Karamchandani; Piyush Nagasubramaniam; Sencun Zhu; Dinghao Wu
2026-07-06
2026
arXiv
arXiv
Preprint. 10 pages, 2 figures, 4 tables
cs.CR
arxiv-api
2607.05029
2026-07-29T08:06:54
ale-0418
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Distributed Attacks in Persistent-State AI Control
https://arxiv.org/abs/2607.02514
external
arxiv.org
Extends AI-control evaluation to coding agents shipping code that persists across sessions, showing a misaligned agent can spread an attack across successive PRs to evade per-transcript monitors, and adds a stateful link-tracker monitor that cuts evasion from 93% to 47%.
Extends AI-control evaluation to coding agents shipping code that persists across sessions, showing a misaligned agent can spread an attack across successive PRs to evade per-transcript monitors, and adds a stateful link-tracker monitor that cuts evasion from 93% to 47%.
Extends AI-control evaluation to coding agents shipping code that persists across sessions, showing a misaligned agent can spread an attack across successive PRs to evade per-transcript monitors, and adds a stateful link-tracker monitor that cuts evasion from 93% to 47%.
Evaluation data is used as the feedback signal for improving loop behavior. Extends AI-control evaluation to coding agents shipping code that persists across sessions, showing a misaligned agent can spread an attack across successive PRs to evade per-transcript monitors, and adds a stateful link-tracker monitor that cu...
Use Distributed Attacks in Persistent-State AI Control to bound risk before recurring or unattended execution.
Research source arXiv:2607.02514; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,073
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1073
Govern
govern
Bound permissions, cost, failure, and escalation.
verification;state
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.02514
[2607.02514] Distributed Attacks in Persistent-State AI Control
As AI coding agents become more autonomous, they increasingly ship code iteratively, with the codebase persisting across sessions. This persistence creates a new attack surface: a misaligned or prompt-injected agent can distribute attacks across pull requests (PRs) and time its payload for the PR with the best natural ...
Josh Hills; Ida Caspary; Asa Cooper Stickland
2026-07-02
2026
arXiv
arXiv
cs.AI
arxiv-api
2607.02514
2026-07-29T08:06:54
ale-0419
Securing Unattended Loops
securing-unattended-loops
Paper
📄
ElephantAgent: Contextual State Continuity in Agentic Systems
https://arxiv.org/abs/2607.01919
external
arxiv.org
Verification protocol that recomputes state digests before each query and logs authorized changes to a trusted-hardware ledger, so an agent's persistent memory and tool descriptions cannot be covertly poisoned between runs and can be rolled back to the last verified state.
Verification protocol that recomputes state digests before each query and logs authorized changes to a trusted-hardware ledger, so an agent's persistent memory and tool descriptions cannot be covertly poisoned between runs and can be rolled back to the last verified state.
Verification protocol that recomputes state digests before each query and logs authorized changes to a trusted-hardware ledger, so an agent's persistent memory and tool descriptions cannot be covertly poisoned between runs and can be rolled back to the last verified state.
Verification is promoted from a final check to a loop-control signal. Verification protocol that recomputes state digests before each query and logs authorized changes to a trusted-hardware ledger, so an agent's persistent memory and tool descriptions cannot be covertly poisoned between runs and can be rolled back to t...
Use ElephantAgent: Contextual State Continuity in Agentic Systems to bound risk before recurring or unattended execution.
Research source arXiv:2607.01919; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,074
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1074
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;context;verification;state
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.01919
[2607.01919] ElephantAgent: Contextual State Continuity in Agentic Systems
Agentic systems enhance their capabilities by invoking external tools and maintaining persistent memory. However, these external dependencies introduce novel attack surfaces. Recent tool and memory poisoning attacks show that maliciously crafted tool descriptors and poisoned memory can covertly bias agent behavior. The...
Jiankai Jin; Xiangzheng Zhang; Zhao Liu; Wenzhuo Xu; Dongdong Yang; Deyue Zhang; Quanchen Zou
2026-07-02
2026
arXiv
arXiv
cs.AI
arxiv-api
2607.01919
2026-07-29T08:06:54
ale-0420
Securing Unattended Loops
securing-unattended-loops
Tool
🧰
Cloudflare security-audit-skill
https://github.com/cloudflare/security-audit-skill
external
github.com
Cloudflare's open-sourced six-phase audit pipeline in which separate validation agents try to disprove each finding and fresh agents independently verify every claim against source code, emitting schema-validated findings that accumulate across repeated runs.
Cloudflare's open-sourced six-phase audit pipeline in which separate validation agents try to disprove each finding and fresh agents independently verify every claim against source code, emitting schema-validated findings that accumulate across repeated runs.
Cloudflare's open-sourced six-phase audit pipeline in which separate validation agents try to disprove each finding and fresh agents independently verify every claim against source code, emitting schema-validated findings that accumulate across repeated runs.
The contribution is machine-readable and validation-friendly. Cloudflare's open-sourced six-phase audit pipeline in which separate validation agents try to disprove each finding and fresh agents independently verify every claim against source code, emitting schema-validated findings that accumulate across repeated runs...
Use Cloudflare security-audit-skill to bound risk before recurring or unattended execution.
Inspectable GitHub source (2,685 stars; 196 forks; MIT license; updated 2026-07-29); popularity is context, not proof of reliability.
medium
README.md
1,075
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1075
Govern
govern
Bound permissions, cost, failure, and escalation.
verification
builder;operator;security
operations
enabling
source-implementation
A
ok
https://github.com/cloudflare/security-audit-skill
GitHub - cloudflare/security-audit-skill: A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings · GitHub
A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings - cloudflare/security-audit-skill
2026-06-18
2026
cloudflare/security-audit-skill
GitHub
github-api
cloudflare/security-audit-skill
2685
196
MIT
2026-06-18T14:08:44Z
2026-07-29T07:38:44Z
2026-07-29T08:06:54
ale-0421
Securing Unattended Loops
securing-unattended-loops
Paper
📄
The Balkanization of Execution-Security Research for AI Coding Agents
https://arxiv.org/abs/2607.05743
external
arxiv.org
Systematizes 39 papers on the execution layer around coding agents, spanning sandbox isolation, capability control, TOCTOU races, MCP threats, and egress control, surfacing five cross-cutting gaps and four verified CVEs in production agent harnesses.
Systematizes 39 papers on the execution layer around coding agents, spanning sandbox isolation, capability control, TOCTOU races, MCP threats, and egress control, surfacing five cross-cutting gaps and four verified CVEs in production agent harnesses.
Systematizes 39 papers on the execution layer around coding agents, spanning sandbox isolation, capability control, TOCTOU races, MCP threats, and egress control, surfacing five cross-cutting gaps and four verified CVEs in production agent harnesses.
Verification is promoted from a final check to a loop-control signal. Systematizes 39 papers on the execution layer around coding agents, spanning sandbox isolation, capability control, TOCTOU races, MCP threats, and egress control, surfacing five cross-cutting gaps and four verified CVEs in production agent harnesses.
Use The Balkanization of Execution-Security Research for AI Coding Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.05743; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,076
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1076
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;verification
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.05743
[2607.05743] The Balkanization of Execution-Security Research for AI Coding Agents: Isolation, Access Control, and Time-of-Check-to-Time-of-Use Vulnerabilities
AI coding agents now read repositories, call tools, and execute shell commands with limited human oversight, and a fast-growing body of work studies whether the execution layer around them is actually safe. That literature is scattered. Papers on sandbox isolation, capability and access control, policy enforcement, tim...
Mohammadreza Rashidi
2026-07-07
2026
arXiv
arXiv
18 pages, 15 figures, 6 tables. Systematizes 39 execution-security papers (2023-2026) into 17 verified categories. Machine-readable corpus and verification script released as a supplementary artifact
cs.CR
arxiv-api
2607.05743
2026-07-29T08:06:54
ale-0422
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Context-to-Execution Integrity for LLM Agents
https://arxiv.org/abs/2607.06000
external
arxiv.org
Execution-boundary system where a deterministic gate admits a tool call only after field authority, exact-effect authorization, and invocation authority all bind to the same action manifest, protecting loops that read attacker-writable context.
Execution-boundary system where a deterministic gate admits a tool call only after field authority, exact-effect authorization, and invocation authority all bind to the same action manifest, protecting loops that read attacker-writable context.
Execution-boundary system where a deterministic gate admits a tool call only after field authority, exact-effect authorization, and invocation authority all bind to the same action manifest, protecting loops that read attacker-writable context.
Context is managed as durable loop state rather than a single prompt payload. Execution-boundary system where a deterministic gate admits a tool call only after field authority, exact-effect authorization, and invocation authority all bind to the same action manifest, protecting loops that read attacker-writable contex...
Use Context-to-Execution Integrity for LLM Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.06000; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,077
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1077
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;context
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.06000
[2607.06000] Context-to-Execution Integrity for LLM Agents
Language-model agents read attacker-writable context to solve tasks. Tool execution needs a separate authority check for protected sink fields, sink-interpreted payloads, and the invocation event. Context-to-Execution Integrity (CXI) is an execution-boundary system for this setting. Policies mark protected sink fields,...
Igor Santos-Grueiro
2026-07-07
2026
arXiv
arXiv
20 pages
cs.CR
arxiv-api
2607.06000
2026-07-29T08:06:54
ale-0423
Securing Unattended Loops
securing-unattended-loops
Paper
📄
When Agents Remember Too Much: Memory Poisoning Attacks on Large Language Model Agents
https://arxiv.org/abs/2607.06595
external
arxiv.org
GhostWriter is a two-phase attack that poisons the long-term memory store of tool-using personal agents so injected content persists across runs and activates in later tasks.
GhostWriter is a two-phase attack that poisons the long-term memory store of tool-using personal agents so injected content persists across runs and activates in later tasks.
GhostWriter is a two-phase attack that poisons the long-term memory store of tool-using personal agents so injected content persists across runs and activates in later tasks.
Persistent memory is treated as an external runtime artifact. GhostWriter is a two-phase attack that poisons the long-term memory store of tool-using personal agents so injected content persists across runs and activates in later tasks.
Use When Agents Remember Too Much: Memory Poisoning Attacks on Large Language Model Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.06595; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,078
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1078
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;context;state
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.06595
[2607.06595] When Agents Remember Too Much: Memory Poisoning Attacks on Large Language Model Agents
Personal AI agents powered by large language models can reason and act using available tools to access emails, manage calendars, and push code to remote repositories, all with minimal oversight. When augmented with long-term memory, an agent can recall specific details relevant to the current task, reducing the need fo...
George Torres; Sharad Shrestha; Satyajayant Misra
2026-07-06
2026
arXiv
arXiv
cs.CR
arxiv-api
2607.06595
2026-07-29T08:06:54
ale-0424
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Token-Flow Firewall: Semantic Runtime Auditing for Persistent AI Agents
https://arxiv.org/abs/2607.08395
external
arxiv.org
Proposes TokenWall, a runtime firewall that audits a long-lived agent's semantic flows (memory updates, tool arguments, inter-component messages) before they reach privileged sinks, reporting attack success reduced to 12.5% with a 97.4% benign pass rate and 0.69s added latency.
Proposes TokenWall, a runtime firewall that audits a long-lived agent's semantic flows (memory updates, tool arguments, inter-component messages) before they reach privileged sinks, reporting attack success reduced to 12.5% with a 97.4% benign pass rate and 0.69s added latency.
Proposes TokenWall, a runtime firewall that audits a long-lived agent's semantic flows (memory updates, tool arguments, inter-component messages) before they reach privileged sinks, reporting attack success reduced to 12.5% with a 97.4% benign pass rate and 0.69s added latency.
Persistent memory is treated as an external runtime artifact. Proposes TokenWall, a runtime firewall that audits a long-lived agent's semantic flows (memory updates, tool arguments, inter-component messages) before they reach privileged sinks, reporting attack success reduced to 12.5% with a 97.4% benign pass rate and ...
Use Token-Flow Firewall: Semantic Runtime Auditing for Persistent AI Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.08395; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,079
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1079
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;context;state;budget
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.08395
[2607.08395] Token-Flow Firewall: Semantic Runtime Auditing for Persistent AI Agents
Persistent AI agents extend large language models (LLMs) beyond single-turn interaction into long-lived software systems. Unlike traditional chat assistants, unsafe content in these agents can propagate through persistent state, reusable skills, and tool-mediated interactions, creating a substantially larger semantic a...
Puji Wang; Yingchen Zhang; Ruqing Zhang; Jiafeng Guo; Xueqi Cheng
2026-07-09
2026
arXiv
arXiv
cs.CR
arxiv-api
2607.08395
2026-07-29T08:06:54
ale-0425
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Prismata: Confining Cross-Site Prompt Injection in Web Agents
https://arxiv.org/abs/2607.08147
external
arxiv.org
Applies contextual least privilege to web agents by dynamically labeling page content with trust levels and mechanically confining what the agent can see and do, cutting cross-site prompt-injection attack success on benign pages without requiring developer annotations.
Applies contextual least privilege to web agents by dynamically labeling page content with trust levels and mechanically confining what the agent can see and do, cutting cross-site prompt-injection attack success on benign pages without requiring developer annotations.
Applies contextual least privilege to web agents by dynamically labeling page content with trust levels and mechanically confining what the agent can see and do, cutting cross-site prompt-injection attack success on benign pages without requiring developer annotations.
Untrusted intake is treated as a loop-level security boundary. Applies contextual least privilege to web agents by dynamically labeling page content with trust levels and mechanically confining what the agent can see and do, cutting cross-site prompt-injection attack success on benign pages without requiring developer ...
Use Prismata: Confining Cross-Site Prompt Injection in Web Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.08147; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,080
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1080
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;budget;escalation
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.08147
[2607.08147] Prismata: Confining Cross-Site Prompt Injection in Web Agents
Autonomous web agents promise to automate everyday browsing tasks, but inherit one of the web's oldest attack surfaces. Cross-Site Scripting proved that mixing trusted and untrusted content is dangerous, even on benign pages. Agents resurface this risk by interpreting natural language as instructions, allowing third-pa...
Corban Villa; Alp Eren Ozdarendeli; Sijun Tan; Raluca Ada Popa
2026-07-09
2026
arXiv
arXiv
cs.CR
arxiv-api
2607.08147
2026-07-29T08:06:54
ale-0426
Securing Unattended Loops
securing-unattended-loops
Paper
📄
TRACE: A Two-Channel Robust Attribution Watermark via Complementary Embeddings for LLM-Agent Trajectories
https://arxiv.org/abs/2607.08400
external
arxiv.org
Embeds a two-channel attribution watermark in LLM-agent trajectory logs (one channel keyed on content for deletion resistance, one on log structure for rewrite resistance) so provenance survives an adversary with full read/write access, reporting detection scores near z = 100 on long-horizon ToolBench and ALFWorld traj...
Embeds a two-channel attribution watermark in LLM-agent trajectory logs (one channel keyed on content for deletion resistance, one on log structure for rewrite resistance) so provenance survives an adversary with full read/write access, reporting detection scores near z = 100 on long-horizon ToolBench and ALFWorld traj...
Embeds a two-channel attribution watermark in LLM-agent trajectory logs (one channel keyed on content for deletion resistance, one on log structure for rewrite resistance) so provenance survives an adversary with full read/write access, reporting detection scores near z = 100 on long-horizon ToolBench and ALFWorld traj...
The work targets tasks that exceed a single context window or prompt session. Embeds a two-channel attribution watermark in LLM-agent trajectory logs (one channel keyed on content for deletion resistance, one on log structure for rewrite resistance) so provenance survives an adversary with full read/write access, repor...
Use TRACE: A Two-Channel Robust Attribution Watermark via Complementary Embeddings for LLM-Agent Trajectories to bound risk before recurring or unattended execution.
Research source arXiv:2607.08400; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,081
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1081
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;budget;escalation
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.08400
[2607.08400] TRACE: A Two-Channel Robust Attribution Watermark via Complementary Embeddings for LLM-Agent Trajectories
LLM agents reach users through resellers, who may rebrand a developer's agent or substitute a cheaper model. When provenance is disputed, attribution rests on the trajectory log (the record of tool calls, observations, and executed actions, not the model's reasoning), which the reseller stores and processes to meter us...
Zheng Gao; Xiaoyu Li; Xiaoyan Feng; Jiaojiao Jiang; Yang Song; Yulei Sui; Zhenchang Xing; Liming Zhu
2026-07-09
2026
arXiv
arXiv
cs.CR
arxiv-api
2607.08400
2026-07-29T08:06:54
ale-0427
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Beyond Attack-Success Rate: Action-Graded Severity Scale for Tool-Using AI Agents
https://arxiv.org/abs/2607.07474
external
arxiv.org
Replaces binary attack-success red-teaming metrics with a seven-level ordinal severity rubric (L0-L6) that grades harm along the agent's tool-call trajectory by action reversibility, scope expansion, and privilege escalation, validated with deterministic analysis and frontier-model judges across multiple victim models ...
Replaces binary attack-success red-teaming metrics with a seven-level ordinal severity rubric (L0-L6) that grades harm along the agent's tool-call trajectory by action reversibility, scope expansion, and privilege escalation, validated with deterministic analysis and frontier-model judges across multiple victim models ...
Replaces binary attack-success red-teaming metrics with a seven-level ordinal severity rubric (L0-L6) that grades harm along the agent's tool-call trajectory by action reversibility, scope expansion, and privilege escalation, validated with deterministic analysis and frontier-model judges across multiple victim models ...
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Replaces binary attack-success red-teaming metrics with a seven-level ordinal severity rubric (L0-L6) that grades harm along the agent's tool-call trajectory by action reversibility, scope expansion, and privilege escalation, vali...
Use Beyond Attack-Success Rate: Action-Graded Severity Scale for Tool-Using AI Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.07474; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,082
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1082
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;escalation
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.07474
[2607.07474] Beyond Attack-Success Rate: Action-Graded Severity Scale for Tool-Using AI Agents
Agentic red-teaming benchmarks report whether an injected agent was compromised as a single bit: the attack succeeded, or it did not. We argue that this binary attack-success rate discards the information a defender most needs, namely how harmful the resulting action was. We introduce an action-graded harm rubric that ...
Harry Owiredu-Ashley
2026-07-08
2026
arXiv
arXiv
8 pages, 6 figures. Code and artifacts: https://github.com/Harry-Ashley/action-graded-severity
cs.CR
arxiv-api
2607.07474
2026-07-29T08:06:54
ale-0428
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Beware of Agentic Botnets: Scalable Untargeted Promptware Attacks via Universal and Transferable Adversarial HalluSquatting
https://arxiv.org/abs/2607.07433
external
arxiv.org
Introduces adversarial hallucination squatting, in which attackers pre-register resource names LLMs predictably hallucinate (at rates up to 85-100%) and plant universal, cross-model-transferable promptware payloads on the open web, reaching agent loops that autonomously ingest internet content with no direct injection ...
Introduces adversarial hallucination squatting, in which attackers pre-register resource names LLMs predictably hallucinate (at rates up to 85-100%) and plant universal, cross-model-transferable promptware payloads on the open web, reaching agent loops that autonomously ingest internet content with no direct injection ...
Introduces adversarial hallucination squatting, in which attackers pre-register resource names LLMs predictably hallucinate (at rates up to 85-100%) and plant universal, cross-model-transferable promptware payloads on the open web, reaching agent loops that autonomously ingest internet content with no direct injection ...
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Introduces adversarial hallucination squatting, in which attackers pre-register resource names LLMs predictably hallucinate (at rates up to 85-100%) and plant universal, cross-model-transferable promptware payloads on the open web...
Use Beware of Agentic Botnets: Scalable Untargeted Promptware Attacks via Universal and Transferable Adversarial HalluSquatting to bound risk before recurring or unattended execution.
Research source arXiv:2607.07433; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,083
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1083
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;budget;escalation
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.07433
[2607.07433] Beware of Agentic Botnets: Scalable Untargeted Promptware Attacks via Universal and Transferable Adversarial HalluSquatting
The growing adoption of agentic LLM applications has introduced a new threat previously named as promptware. While prior work has established that adversaries can exploit direct channels to LLM applications to apply promptware under weak threat models, many applications do not provide any direct channels that could be ...
Aya Spira; Stav Cohen; Elad Feldman; Ron Bitton; Avishai Wool; Ben Nassi
2026-07-08
2026
arXiv
arXiv
Website: https://sites.google.com/view/agentic-botnets/home
cs.CR
arxiv-api
2607.07433
2026-07-29T08:06:54
ale-0429
Securing Unattended Loops
securing-unattended-loops
Blog
📝
GitLost: How We Tricked GitHub's AI Agent into Leaking Private Repos
https://noma.security/blog/gitlost-how-we-tricked-githubs-ai-agent-into-leaking-private-repos/
external
noma.security
Noma Labs researcher Sasi Levi shows how hidden plain-English instructions in a malicious GitHub Issue make an issue-triggered GitHub Agentic Workflows agent exfiltrate private-repo contents into public comments, bypassing GitHub's data-leak guardrails with a one-word reframe, disclosed responsibly to GitHub.
Noma Labs researcher Sasi Levi shows how hidden plain-English instructions in a malicious GitHub Issue make an issue-triggered GitHub Agentic Workflows agent exfiltrate private-repo contents into public comments, bypassing GitHub's data-leak guardrails with a one-word reframe, disclosed responsibly to GitHub.
Noma Labs researcher Sasi Levi shows how hidden plain-English instructions in a malicious GitHub Issue make an issue-triggered GitHub Agentic Workflows agent exfiltrate private-repo contents into public comments, bypassing GitHub's data-leak guardrails with a one-word reframe, disclosed responsibly to GitHub.
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Noma Labs researcher Sasi Levi shows how hidden plain-English instructions in a malicious GitHub Issue make an issue-triggered GitHub Agentic Workflows agent exfiltrate private-repo contents into public comments, bypassing GitHub'...
Use GitLost: How We Tricked GitHub's AI Agent into Leaking Private Repos to bound risk before recurring or unattended execution.
Contextual source from noma.security; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,084
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1084
Govern
govern
Bound permissions, cost, failure, and escalation.
trigger;intake
operator;security
operations
enabling
practitioner-analysis
B
ok
https://noma.security/blog/gitlost-how-we-tricked-githubs-ai-agent-into-leaking-private-repos/
GitLost: How We Tricked GitHub’s AI Agent into Leaking Private Repos - Noma Security
TL;DR: Noma Labs discovered a critical prompt injection vulnerability within GitHub’s new Agentic Workflows, allowing an unauthenticated attacker to silently pull data from private repositories by posting a crafted GitHub Issue in a public repository belonging to the same organization as the private repositories. Noma ...
noma.security
domain-fallback
2026-07-29T08:06:54
ale-0430
Securing Unattended Loops
securing-unattended-loops
Benchmark
🧪
ScopeJudge: Cost-Aware Pre-Execution Gating for Offensive Security Agents
https://arxiv.org/abs/2607.07774
external
arxiv.org
Benchmark of 4,897 tool calls from offensive-security agent trajectories, labeled by professional penetration testers, for studying pre-execution gating where a cheap trusted LLM judge accepts or rejects each proposed call before it runs and the engagement boundary must be inferred from the request rather than a fixed ...
Benchmark of 4,897 tool calls from offensive-security agent trajectories, labeled by professional penetration testers, for studying pre-execution gating where a cheap trusted LLM judge accepts or rejects each proposed call before it runs and the engagement boundary must be inferred from the request rather than a fixed ...
Benchmark of 4,897 tool calls from offensive-security agent trajectories, labeled by professional penetration testers, for studying pre-execution gating where a cheap trusted LLM judge accepts or rejects each proposed call before it runs and the engagement boundary must be inferred from the request rather than a fixed ...
The work turns loop quality into a measurable task or score. Benchmark of 4,897 tool calls from offensive-security agent trajectories, labeled by professional penetration testers, for studying pre-execution gating where a cheap trusted LLM judge accepts or rejects each proposed call before it runs and the engagement bo...
Use ScopeJudge: Cost-Aware Pre-Execution Gating for Offensive Security Agents to bound risk before recurring or unattended execution.
Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope.
high
README.md
1,085
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1085
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;verification;budget
researcher;evaluator;operator;security
operations
enabling
benchmark
A
ok
https://arxiv.org/abs/2607.07774
[2607.07774] ScopeJudge: Cost-Aware Pre-Execution Gating for Offensive Security Agents
As LLM agents take on offensive security work, a single out-of-scope tool call can breach a client's engagement boundary, disrupt production, or void a bug-bounty finding. Unlike a fixed safety policy, the boundary that matters is declared in the user's request and must be inferred from intent. That challenge is sharpe...
Shane Caldwell; Max Harley; Ads Dawson; Michael Kouremetis; Vincent Abruzzo; Will Pearce
2026-07-08
2026
arXiv
arXiv
22 pages, 4 figures, 4 tables
cs.CR
arxiv-api
2607.07774
2026-07-29T08:06:54
ale-0431
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Multi-Agent AI Control: Distributed Attacks Hamper Per-Instance Monitors
https://arxiv.org/abs/2607.07368
external
arxiv.org
Studies AI control when several agents operating jointly on shared infrastructure pursue a malicious goal, evaluated on FakeLab (synthetic AI-lab codebase: 9 services, 86 benign tasks, 4 attack scenarios); finds a fragmentation effect (the more agents coordinate an attack, the less likely per-agent monitors catch any s...
Studies AI control when several agents operating jointly on shared infrastructure pursue a malicious goal, evaluated on FakeLab (synthetic AI-lab codebase: 9 services, 86 benign tasks, 4 attack scenarios); finds a fragmentation effect (the more agents coordinate an attack, the less likely per-agent monitors catch any s...
Studies AI control when several agents operating jointly on shared infrastructure pursue a malicious goal, evaluated on FakeLab (synthetic AI-lab codebase: 9 services, 86 benign tasks, 4 attack scenarios); finds a fragmentation effect (the more agents coordinate an attack, the less likely per-agent monitors catch any s...
The work separates roles across agents, verifiers, or orchestration layers. Studies AI control when several agents operating jointly on shared infrastructure pursue a malicious goal, evaluated on FakeLab (synthetic AI-lab codebase: 9 services, 86 benign tasks, 4 attack scenarios); finds a fragmentation effect (the more...
Use Multi-Agent AI Control: Distributed Attacks Hamper Per-Instance Monitors to bound risk before recurring or unattended execution.
Research source arXiv:2607.07368; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,086
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1086
Govern
govern
Bound permissions, cost, failure, and escalation.
objective;delegation
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.07368
[2607.07368] Multi-Agent AI Control: Distributed Attacks Hamper Per-Instance Monitors
AI control is a family of techniques to prevent an AI with malicious goals from subverting its operator's intent. AI Control usually studies a single agent in one trajectory, but real deployments run many agents over shared infrastructure, and the most severe risks (model-weight exfiltration, training-run poisoning) pl...
Oliver Makins; Orazio Angelini; Zohreh Shams; Mary Phuong
2026-07-08
2026
arXiv
arXiv
Submitted to NeurIPS; 81 pages; 32 figures and 24 tables
cs.LG
arxiv-api
2607.07368
2026-07-29T08:06:54
ale-0432
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Mitigating Taint-Style Vulnerabilities in MCP Servers via Security-Aware Tool Descriptions
https://arxiv.org/abs/2607.07461
external
arxiv.org
Finds taint-style flaws make up a substantial fraction of MCP-server vulnerabilities and normally demand context-specific code fixes, then proposes SPELLSMITH, which hardens tool descriptions with security-aware behavioral guidance so the agent's own self-reflection steers it away from triggering the vulnerable flows, ...
Finds taint-style flaws make up a substantial fraction of MCP-server vulnerabilities and normally demand context-specific code fixes, then proposes SPELLSMITH, which hardens tool descriptions with security-aware behavioral guidance so the agent's own self-reflection steers it away from triggering the vulnerable flows, ...
Finds taint-style flaws make up a substantial fraction of MCP-server vulnerabilities and normally demand context-specific code fixes, then proposes SPELLSMITH, which hardens tool descriptions with security-aware behavioral guidance so the agent's own self-reflection steers it away from triggering the vulnerable flows, ...
Context is managed as durable loop state rather than a single prompt payload. Finds taint-style flaws make up a substantial fraction of MCP-server vulnerabilities and normally demand context-specific code fixes, then proposes SPELLSMITH, which hardens tool descriptions with security-aware behavioral guidance so the age...
Use Mitigating Taint-Style Vulnerabilities in MCP Servers via Security-Aware Tool Descriptions to bound risk before recurring or unattended execution.
Research source arXiv:2607.07461; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,087
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1087
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;context
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.07461
[2607.07461] Mitigating Taint-Style Vulnerabilities in MCP Servers via Security-Aware Tool Descriptions
Large language models (LLMs) are increasingly deployed as autonomous agents that interact with external tools and services via the Model Context Protocol (MCP), a standardized interface for dynamic tool invocation. While MCP simplifies integration, it also expands the attack surface and enables generic exploits across ...
Yang Shi; Jiaheng Fu; Yihe Huang; Ruixiang Wu; Chengyao Sun; Kaifeng Huang
2026-07-08
2026
arXiv
arXiv
cs.CR
arxiv-api
2607.07461
2026-07-29T08:06:54
ale-0433
Securing Unattended Loops
securing-unattended-loops
Blog
📝
Factory Droid Shield 2.0: Learned Secret Detection for Autonomous Commits
https://factory.ai/news/droid-shield-2-0
external
factory.ai
Factory upgrades the verification gate on every autonomous Droid commit with a two-model pipeline flanking the deterministic secret scanner, pairing a high-recall risk model with a precision referee.
Factory upgrades the verification gate on every autonomous Droid commit with a two-model pipeline flanking the deterministic secret scanner, pairing a high-recall risk model with a precision referee.
Factory upgrades the verification gate on every autonomous Droid commit with a two-model pipeline flanking the deterministic secret scanner, pairing a high-recall risk model with a precision referee.
Verification is promoted from a final check to a loop-control signal. Factory upgrades the verification gate on every autonomous Droid commit with a two-model pipeline flanking the deterministic secret scanner, pairing a high-recall risk model with a precision referee.
Use Factory Droid Shield 2.0: Learned Secret Detection for Autonomous Commits to bound risk before recurring or unattended execution.
Contextual source from factory.ai; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,088
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1088
Govern
govern
Bound permissions, cost, failure, and escalation.
verification
operator;security
operations
enabling
practitioner-analysis
B
ok
https://factory.ai/news/droid-shield-2-0
Droid Shield 2.0: learned secret detection | Factory.ai Factory.ai Logo Arrow Right Icon
Droid Shield 2.0: learned secret detection Factory's Droids write, refactor, and commit code autonomously at a volume th...
Factory
2026-07-01
2026
Factory
html-meta
2026-07-29T08:06:54
ale-0434
Securing Unattended Loops
securing-unattended-loops
Tool
🧰
destructive_command_guard
https://github.com/Dicklesworthstone/destructive_command_guard
external
github.com
Rust safety hook that intercepts and blocks destructive Git and shell commands (hard resets, recursive deletes, database drops) before AI coding agents execute them, across Claude Code and other harnesses.
Rust safety hook that intercepts and blocks destructive Git and shell commands (hard resets, recursive deletes, database drops) before AI coding agents execute them, across Claude Code and other harnesses.
Rust safety hook that intercepts and blocks destructive Git and shell commands (hard resets, recursive deletes, database drops) before AI coding agents execute them, across Claude Code and other harnesses.
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Rust safety hook that intercepts and blocks destructive Git and shell commands (hard resets, recursive deletes, database drops) before AI coding agents execute them, across Claude Code and other harnesses.
Use destructive_command_guard to bound risk before recurring or unattended execution.
Inspectable GitHub source (5,448 stars; 211 forks; NOASSERTION license; updated 2026-07-29); popularity is context, not proof of reliability.
medium
README.md
1,089
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1089
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;budget;escalation
builder;operator;security
operations
enabling
source-implementation
A
ok
https://github.com/Dicklesworthstone/destructive_command_guard
GitHub - Dicklesworthstone/destructive_command_guard: The Destructive Command Guard (dcg) is for blocking dangerous git and shell commands from being executed by agents. · GitHub
The Destructive Command Guard (dcg) is for blocking dangerous git and shell commands from being executed by agents. - Dicklesworthstone/destructive_command_guard
2026-01-07
2026
Dicklesworthstone/destructive_command_guard
GitHub
github-api
Dicklesworthstone/destructive_command_guard
5448
211
NOASSERTION
2026-01-07T22:27:34Z
2026-07-29T07:21:02Z
2026-07-29T08:06:54
ale-0435
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Friendly Fire: Hijacking Defensive Cyber AI Agents for Remote Code Execution
https://ainowinstitute.org/publications/friendly-fire-exploit-brief
external
ainowinstitute.org
Proof-of-concept showing prompt injections spread across ordinary repository files can hijack defensive security agents into remote code execution, demonstrating that even security-focused agent loops inherit the untrusted-content attack surface.
Proof-of-concept showing prompt injections spread across ordinary repository files can hijack defensive security agents into remote code execution, demonstrating that even security-focused agent loops inherit the untrusted-content attack surface.
Proof-of-concept showing prompt injections spread across ordinary repository files can hijack defensive security agents into remote code execution, demonstrating that even security-focused agent loops inherit the untrusted-content attack surface.
Untrusted intake is treated as a loop-level security boundary. Proof-of-concept showing prompt injections spread across ordinary repository files can hijack defensive security agents into remote code execution, demonstrating that even security-focused agent loops inherit the untrusted-content attack surface.
Use Friendly Fire: Hijacking Defensive Cyber AI Agents for Remote Code Execution to bound risk before recurring or unattended execution.
Research source; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,090
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1090
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;budget;escalation
researcher;evaluator;operator;security
operations
enabling
research-paper
A
ok
https://ainowinstitute.org/publications/friendly-fire-exploit-brief
Friendly Fire: Hijacking Defensive Cyber AI Agents for Remote Code Execution - AI Now Institute
AI Now’s latest research demonstrates a critical attack vector on popular AI agents, built by Anthropic and OpenAI, when used for defensive purposes that actually turn the agent against its user.
Boyan Milanov
2026-07-08
2026
AI Now Institute
html-meta
2026-07-29T08:06:54
ale-0436
Securing Unattended Loops
securing-unattended-loops
Blog
📝
How We Contain Claude Across Products
https://www.anthropic.com/engineering/how-we-contain-claude
external
www.anthropic.com
Anthropic engineering on capping agent blast radius with three containment architectures matched to threat models, including ephemeral sandbox containers for untrusted code execution.
Anthropic engineering on capping agent blast radius with three containment architectures matched to threat models, including ephemeral sandbox containers for untrusted code execution.
Anthropic engineering on capping agent blast radius with three containment architectures matched to threat models, including ephemeral sandbox containers for untrusted code execution.
Execution isolation and permission boundaries are part of the design. Anthropic engineering on capping agent blast radius with three containment architectures matched to threat models, including ephemeral sandbox containers for untrusted code execution.
Use How We Contain Claude Across Products to bound risk before recurring or unattended execution.
Contextual source from www.anthropic.com; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,091
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1091
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace
operator;security
operations
enabling
practitioner-analysis
B
ok
https://www.anthropic.com/engineering/how-we-contain-claude
How we contain Claude across products \ Anthropic
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
Anthropic
domain-fallback
2026-07-29T08:06:54
ale-0437
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Rethinking MCP Security: A Large-Scale Study of Runtime MCP Servers and Scanner Reliability
https://arxiv.org/abs/2607.11086
external
arxiv.org
Large-scale study of live MCP servers finding widespread security weaknesses and that existing MCP security scanners miss or misreport many of them, a gap for anyone gating agent tool access on scanner output.
Large-scale study of live MCP servers finding widespread security weaknesses and that existing MCP security scanners miss or misreport many of them, a gap for anyone gating agent tool access on scanner output.
Large-scale study of live MCP servers finding widespread security weaknesses and that existing MCP security scanners miss or misreport many of them, a gap for anyone gating agent tool access on scanner output.
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Large-scale study of live MCP servers finding widespread security weaknesses and that existing MCP security scanners miss or misreport many of them, a gap for anyone gating agent tool access on scanner output.
Use Rethinking MCP Security: A Large-Scale Study of Runtime MCP Servers and Scanner Reliability to bound risk before recurring or unattended execution.
Research source arXiv:2607.11086; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,092
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1092
2026-07-15
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.11086
[2607.11086] Rethinking MCP Security: A Large-Scale Study of Runtime MCP Servers and Security Scanner Reliability
The Model Context Protocol (MCP) has rapidly established itself as a standard interface for enabling LLM-based agents to interact with external tools and services. As MCP servers are increasingly entrusted with security-sensitive operations, understanding their real-world risks has become critical. In practice, due to ...
Pei Chen; Baichao An; Mengying Wu; Binwang Wan; Geng Hong; Jinsong Chen; Xudong Pan; Jiarun Dai; Min Yang
2026-07-13
2026
arXiv
arXiv
18 pages, 11 figures, and 10 tables. This article substantially extends the preliminary 3-page MCPZoo dataset release arXiv:2512.15144. Includes appendices
cs.CR
arxiv-api
2607.11086
2026-07-29T08:06:54
ale-0438
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Agent Hacks Agent: Autoresearch for Production-Agent Red-Teaming
https://arxiv.org/abs/2607.11698
external
arxiv.org
Uses an autonomous research loop to red-team production agents, having one agent iteratively discover, reproduce, and refine attacks against another, turning red-teaming itself into a recurring verified loop.
Uses an autonomous research loop to red-team production agents, having one agent iteratively discover, reproduce, and refine attacks against another, turning red-teaming itself into a recurring verified loop.
Uses an autonomous research loop to red-team production agents, having one agent iteratively discover, reproduce, and refine attacks against another, turning red-teaming itself into a recurring verified loop.
Verification is promoted from a final check to a loop-control signal. Uses an autonomous research loop to red-team production agents, having one agent iteratively discover, reproduce, and refine attacks against another, turning red-teaming itself into a recurring verified loop.
Use Agent Hacks Agent: Autoresearch for Production-Agent Red-Teaming to bound risk before recurring or unattended execution.
Research source arXiv:2607.11698; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,093
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1093
2026-07-15
Govern
govern
Bound permissions, cost, failure, and escalation.
intake;verification
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.11698
[2607.11698] Agent Hacks Agent: Autoresearch for Production-Agent Red-Teaming
Production LLM agents such as Claude Code and Codex operate over untrusted content, files, commands, and workspace state, making safety failures directly actionable. Red-teaming must therefore keep pace with evolving models and tools. Existing approaches mainly optimize attack success and preserve artifacts such as ben...
Xutao Mao; Xiang Zheng; Cong Wang
2026-07-13
2026
arXiv
arXiv
cs.CR
arxiv-api
2607.11698
2026-07-29T08:06:54
ale-0439
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Temporary Authority, Permanent Effects: Commit-Time Authorization for LLM Agents
https://arxiv.org/abs/2607.10487
external
arxiv.org
Proposes binding an agent's authority to the moment of commit rather than the moment of request, so a permission granted mid-run cannot be replayed later to cause irreversible effects in unattended execution.
Proposes binding an agent's authority to the moment of commit rather than the moment of request, so a permission granted mid-run cannot be replayed later to cause irreversible effects in unattended execution.
Proposes binding an agent's authority to the moment of commit rather than the moment of request, so a permission granted mid-run cannot be replayed later to cause irreversible effects in unattended execution.
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Proposes binding an agent's authority to the moment of commit rather than the moment of request, so a permission granted mid-run cannot be replayed later to cause irreversible effects in unattended execution.
Use Temporary Authority, Permanent Effects: Commit-Time Authorization for LLM Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.10487; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,094
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1094
2026-07-15
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.10487
[2607.10487] Temporary Authority, Permanent Effects: Commit-Time Authorization for LLM Agents
LLM agents can commit durable effects from authority evidence that was valid earlier in execution: a DOM snapshot, approval epoch, version witness, branch token, or worker result. We study the commit boundary at which earlier authority evidence no longer authorizes a durable effect. We call this property commit-time au...
Igor Santos-Grueiro
2026-07-11
2026
arXiv
arXiv
20 pages
cs.CR
arxiv-api
2607.10487
2026-07-29T08:06:54
ale-0440
Securing Unattended Loops
securing-unattended-loops
Paper
📄
ANCHOR: Automated Alignment Auditing for CLI Agents on Real-World Harm
https://arxiv.org/abs/2607.10455
external
arxiv.org
Automated auditing framework that probes CLI coding agents for real-world harmful behavior and grades alignment, giving unattended-agent operators a repeatable safety check rather than manual spot review.
Automated auditing framework that probes CLI coding agents for real-world harmful behavior and grades alignment, giving unattended-agent operators a repeatable safety check rather than manual spot review.
Automated auditing framework that probes CLI coding agents for real-world harmful behavior and grades alignment, giving unattended-agent operators a repeatable safety check rather than manual spot review.
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Automated auditing framework that probes CLI coding agents for real-world harmful behavior and grades alignment, giving unattended-agent operators a repeatable safety check rather than manual spot review.
Use ANCHOR: Automated Alignment Auditing for CLI Agents on Real-World Harm to bound risk before recurring or unattended execution.
Research source arXiv:2607.10455; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,095
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1095
2026-07-15
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;budget;escalation
researcher;evaluator;operator;security
operations
enabling
research-paper
A
ok
https://openreview.net/forum?id=YqTodSrPPB
[2607.10455] ANCHOR: Automated Alignment Auditing for CLI Agents on Real-World Harm
Autonomous CLI agents can now execute hundreds of actions across multi-hour sessions: writing code, executing shell commands, browsing the web, and managing cloud infrastructure, all with minimal human oversight. Does greater autonomy invite greater risk? We introduce ANCHOR, an automated auditing framework that stress...
Kefan Song; Yanjun Qi
2026-04-30
2026
Proceedings of the 43rd International Conference on Machine Learning (ICML)
PMLR
Accepted at Proceedings of the 43rd International Conference on Machine Learning (ICML); the linked arXiv record is the available paper version.
cs.AI
ICML OpenReview record
2607.10455
2026-07-29T08:06:54
ale-0441
Securing Unattended Loops
securing-unattended-loops
Tool
🧰
Clawk
https://github.com/clawkwork/clawk
external
github.com
Runs coding agents inside disposable, network-restricted Linux VMs so an unattended or untrusted agent's blast radius is confined to a throwaway sandbox.
Runs coding agents inside disposable, network-restricted Linux VMs so an unattended or untrusted agent's blast radius is confined to a throwaway sandbox.
Runs coding agents inside disposable, network-restricted Linux VMs so an unattended or untrusted agent's blast radius is confined to a throwaway sandbox.
Execution isolation and permission boundaries are part of the design. Runs coding agents inside disposable, network-restricted Linux VMs so an unattended or untrusted agent's blast radius is confined to a throwaway sandbox.
Use Clawk to bound risk before recurring or unattended execution.
Inspectable GitHub source (838 stars; 27 forks; Apache-2.0 license; updated 2026-07-29); popularity is context, not proof of reliability.
medium
README.md
1,096
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1096
2026-07-15
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace
builder;operator;security
operations
enabling
source-implementation
A
ok
https://github.com/clawkwork/clawk
GitHub - clawkwork/clawk: Give coding agents a disposable Linux VM, not your laptop · GitHub
Give coding agents a disposable Linux VM, not your laptop - clawkwork/clawk
2026-07-06
2026
clawkwork/clawk
GitHub
github-api
clawkwork/clawk
838
27
Apache-2.0
2026-07-06T21:30:38Z
2026-07-29T07:14:40Z
2026-07-29T08:06:54
ale-0442
Securing Unattended Loops
securing-unattended-loops
Blog
📝
Auto-Review of Agent Actions Without Synchronous Human Oversight
https://alignment.openai.com/auto-review/
external
alignment.openai.com
OpenAI's alignment team on reviewing agent actions asynchronously with automated reviewers when a human cannot watch every step, so oversight scales with agent throughput instead of gating it.
OpenAI's alignment team on reviewing agent actions asynchronously with automated reviewers when a human cannot watch every step, so oversight scales with agent throughput instead of gating it.
OpenAI's alignment team on reviewing agent actions asynchronously with automated reviewers when a human cannot watch every step, so oversight scales with agent throughput instead of gating it.
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. OpenAI's alignment team on reviewing agent actions asynchronously with automated reviewers when a human cannot watch every step, so oversight scales with agent throughput instead of gating it.
Use Auto-Review of Agent Actions Without Synchronous Human Oversight to bound risk before recurring or unattended execution.
Contextual source from alignment.openai.com; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,097
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1097
2026-07-15
Govern
govern
Bound permissions, cost, failure, and escalation.
escalation
operator;security
operations
enabling
practitioner-analysis
B
ok
https://alignment.openai.com/auto-review/
Auto-review of agent actions without synchronous human oversight Illustrative deployment snapshot showing 10,000 actions, most inside the sandbox, and out-of-sandbox actions reviewed by Auto-review.
Auto-review offers a safer default for deploying coding agents, using a separate agent to approve or deny boundary-crossing actions.
Maja Trębacz; Sam Arnesen; Ollie Matthews; Dylan Hurd; Won Park; Owen Lin; Joe Gershenson
2026-04-30
2026
OpenAI
OpenAI
html-meta
2026-07-29T08:06:54
ale-0443
Securing Unattended Loops
securing-unattended-loops
Paper
📄
SAFETY SENTRY: Context-Aware Human Intervention via EXECUTE-ASK-REFUSE Routing
https://arxiv.org/abs/2607.13594
external
arxiv.org
Routes each proposed action among execute, ask a human, and refuse, with one threshold controlling the deployment's risk posture; experiments report stronger overall accuracy and safety recall than the compared baselines.
Routes each proposed action among execute, ask a human, and refuse, with one threshold controlling the deployment's risk posture; experiments report stronger overall accuracy and safety recall than the compared baselines.
Routes each proposed action among execute, ask a human, and refuse, with one threshold controlling the deployment's risk posture; experiments report stronger overall accuracy and safety recall than the compared baselines.
Context is managed as durable loop state rather than a single prompt payload. Routes each proposed action among execute, ask a human, and refuse, with one threshold controlling the deployment's risk posture; experiments report stronger overall accuracy and safety recall than the compared baselines.
Use SAFETY SENTRY: Context-Aware Human Intervention via EXECUTE-ASK-REFUSE Routing to bound risk before recurring or unattended execution.
Research source arXiv:2607.13594; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,098
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1098
2026-07-17
Govern
govern
Bound permissions, cost, failure, and escalation.
context;escalation
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.13594
[2607.13594] SAFETY SENTRY: Context-Aware Human Intervention via EXECUTE-ASK-REFUSE Routing
LLM agents act on real-world environments through tool calls, and a single misjudged action can cause irreversible harm. The standard safeguard is a guard model that labels each proposed action as safe or unsafe, but this binary view conflates two distinct decisions: whether the action is harmful in itself, and whether...
Tianyu Chen; Chujia Hu; Wenjie Wang
2026-07-15
2026
arXiv
arXiv
cs.AI
arxiv-api
2607.13594
2026-07-29T08:06:54
ale-0444
Securing Unattended Loops
securing-unattended-loops
Paper
📄
CAVA: Canonical Action Verification and Attestation for Runtime Governance of Agentic AI Systems
https://arxiv.org/abs/2607.13716
external
arxiv.org
Normalizes heterogeneous tool calls into a canonical runtime action object that can be verified and attested before execution; evaluation spans 96 seeds and 384 variants covering approval binding, tampering, and runtime portability.
Normalizes heterogeneous tool calls into a canonical runtime action object that can be verified and attested before execution; evaluation spans 96 seeds and 384 variants covering approval binding, tampering, and runtime portability.
Normalizes heterogeneous tool calls into a canonical runtime action object that can be verified and attested before execution; evaluation spans 96 seeds and 384 variants covering approval binding, tampering, and runtime portability.
Verification is promoted from a final check to a loop-control signal. Normalizes heterogeneous tool calls into a canonical runtime action object that can be verified and attested before execution; evaluation spans 96 seeds and 384 variants covering approval binding, tampering, and runtime portability.
Use CAVA: Canonical Action Verification and Attestation for Runtime Governance of Agentic AI Systems to bound risk before recurring or unattended execution.
Research source arXiv:2607.13716; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,099
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1099
2026-07-17
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;verification;escalation
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.13716
[2607.13716] CAVA: Canonical Action Verification and Attestation for Runtime Governance of Agentic AI Systems
Agentic AI systems increasingly act through heterogeneous runtimes: local coding hooks, SDK tools, browser automation, managed-agent traces, API gateways, and workflow engines. A single operational act such as publishing code, changing identity state, moving money, or exporting data may therefore be represented by many...
Zexun Wang
2026-07-15
2026
arXiv
arXiv
35 pages. Working paper on canonical action verification, runtime governance, semantic pattern detection, and approval-bound action receipts
cs.AI
arxiv-api
2607.13716
2026-07-29T08:06:54
ale-0445
Securing Unattended Loops
securing-unattended-loops
Paper
📄
How Agents Ask for Permission: User Permissions for AI Agents, from Interfaces to Enforcement
https://arxiv.org/abs/2607.13718
external
arxiv.org
Surveys 21 permission proposals and compares five commercial agents, producing a taxonomy that connects what users see in permission interfaces to how authority is represented and enforced at runtime.
Surveys 21 permission proposals and compares five commercial agents, producing a taxonomy that connects what users see in permission interfaces to how authority is represented and enforced at runtime.
Surveys 21 permission proposals and compares five commercial agents, producing a taxonomy that connects what users see in permission interfaces to how authority is represented and enforced at runtime.
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Surveys 21 permission proposals and compares five commercial agents, producing a taxonomy that connects what users see in permission interfaces to how authority is represented and enforced at runtime.
Use How Agents Ask for Permission: User Permissions for AI Agents, from Interfaces to Enforcement to bound risk before recurring or unattended execution.
Research source arXiv:2607.13718; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,100
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1100
2026-07-17
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.13718
[2607.13718] How Agents Ask for Permission: User Permissions for AI Agents, from Interfaces to Enforcement
As AI agents gain prevalence, users are increasingly exposed to the risks such systems entail. Prompt injection attacks, as well as hallucination, can cause agents to leak private information to third parties. As autonomous systems, agents also present the more active danger of performing sensitive tasks, such as bank ...
Alexandra E. Michael; Franziska Roesner
2026-07-15
2026
arXiv
arXiv
15 pages, 4 figures
cs.CR
arxiv-api
2607.13718
2026-07-29T08:06:54
ale-0446
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Stop Means Stop: Measuring and Repairing the Enforcement Gap in Agent-Framework Control Primitives
https://arxiv.org/abs/2607.14166
external
arxiv.org
Tests six open-source frameworks and finds their stop or approval controls do not behave as execution barriers; 215 of 1,200 live runs performed a side effect during an approval pause, while the proposed admission gate blocks all measured violations with roughly 1 ms overhead.
Tests six open-source frameworks and finds their stop or approval controls do not behave as execution barriers; 215 of 1,200 live runs performed a side effect during an approval pause, while the proposed admission gate blocks all measured violations with roughly 1 ms overhead.
Tests six open-source frameworks and finds their stop or approval controls do not behave as execution barriers; 215 of 1,200 live runs performed a side effect during an approval pause, while the proposed admission gate blocks all measured violations with roughly 1 ms overhead.
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Tests six open-source frameworks and finds their stop or approval controls do not behave as execution barriers; 215 of 1,200 live runs performed a side effect during an approval pause, while the proposed admission gate blocks all ...
Use Stop Means Stop: Measuring and Repairing the Enforcement Gap in Agent-Framework Control Primitives to bound risk before recurring or unattended execution.
Research source arXiv:2607.14166; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,101
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1101
2026-07-17
Govern
govern
Bound permissions, cost, failure, and escalation.
verification;escalation;exit
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.14166
[2607.14166] Stop Means Stop: Measuring and Repairing the Enforcement Gap in Agent-Framework Control Primitives
Production LLM-agent frameworks ship control primitives -- human-in-the-loop approval gates, run cancellation, and execution timeouts -- whose names and documentation imply barrier semantics: while a run is paused, cancelled, or timed out, no gated side effect executes. This contract holds on none of six widely used op...
Sajjad Khan
2026-07-15
2026
arXiv
arXiv
32 pages, 3 figures, 11 tables. Under review at the Journal of Systems and Software. Code: pip install soundgate (PyPI)
cs.SE
arxiv-api
2607.14166
2026-07-29T08:06:54
ale-0447
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Bad Memory: Evaluating Prompt Injection Risks from Memory in Agentic Systems
https://arxiv.org/abs/2607.14611
external
arxiv.org
Evaluates planted memory payloads in Claude Code and Codex across four models, showing that malicious state can affect both current and future sessions and can persist differently across harnesses.
Evaluates planted memory payloads in Claude Code and Codex across four models, showing that malicious state can affect both current and future sessions and can persist differently across harnesses.
Evaluates planted memory payloads in Claude Code and Codex across four models, showing that malicious state can affect both current and future sessions and can persist differently across harnesses.
Persistent memory is treated as an external runtime artifact. Evaluates planted memory payloads in Claude Code and Codex across four models, showing that malicious state can affect both current and future sessions and can persist differently across harnesses.
Use Bad Memory: Evaluating Prompt Injection Risks from Memory in Agentic Systems to bound risk before recurring or unattended execution.
Research source arXiv:2607.14611; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,102
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1102
2026-07-17
Govern
govern
Bound permissions, cost, failure, and escalation.
context;state
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.14611
[2607.14611] Bad Memory: Evaluating Prompt Injection Risks from Memory in Agentic Systems
A growing class of agentic systems maintain persistent state across sessions through memory files, behavioral preferences, and knowledge bases. While this makes agents more useful and self-improving, it also creates a new attack surface for prompt injections in which malicious instructions can be embedded within persis...
Soham Gadgil; David Alexander; Sai Sunku; Franziska Roesner
2026-07-16
2026
arXiv
arXiv
Preprint
cs.CR
arxiv-api
2607.14611
2026-07-29T08:06:54
ale-0448
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Setup Complete, Now You Are Compromised: Weaponizing Setup Instructions Against AI Coding Agents
https://arxiv.org/abs/2607.15143
external
arxiv.org
Demonstrates five setup-instruction attack classes in 12 scenarios across production coding harnesses, including README and dependency attacks; a deterministic pre-install check closes most of the measured gap.
Demonstrates five setup-instruction attack classes in 12 scenarios across production coding harnesses, including README and dependency attacks; a deterministic pre-install check closes most of the measured gap.
Demonstrates five setup-instruction attack classes in 12 scenarios across production coding harnesses, including README and dependency attacks; a deterministic pre-install check closes most of the measured gap.
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Demonstrates five setup-instruction attack classes in 12 scenarios across production coding harnesses, including README and dependency attacks; a deterministic pre-install check closes most of the measured gap.
Use Setup Complete, Now You Are Compromised: Weaponizing Setup Instructions Against AI Coding Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.15143; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,103
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1103
2026-07-17
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;budget;escalation
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.15143
[2607.15143] Setup Complete, Now You Are Compromised: Weaponizing Setup Instructions Against AI Coding Agents
AI coding agents set up projects by reading documentation and installing the dependencies it lists, without verifying their names, sources, or known vulnerabilities. By editing only a README, requirements file, or Makefile, an attacker can redirect the agent to an untrusted registry, a known-vulnerable version, or a wr...
Aadesh Bagmar; Pushkar Saraf
2026-07-16
2026
arXiv
arXiv
cs.CR
arxiv-api
2607.15143
2026-07-29T08:06:54
ale-0449
Securing Unattended Loops
securing-unattended-loops
Paper
📄
SeerGuard: A Safety Framework for Mobile GUI Agents via World Model Prediction
https://arxiv.org/abs/2607.15550
external
arxiv.org
Screens instructions and predicts the consequences of proposed GUI actions before execution with a safety-augmented world model; on Qwen3-VL-8B-Instruct, the reported safety-utility score rises from 0.191 to 0.596 while risk cost falls from 0.347 to 0.130 under the stated settings.
Screens instructions and predicts the consequences of proposed GUI actions before execution with a safety-augmented world model; on Qwen3-VL-8B-Instruct, the reported safety-utility score rises from 0.191 to 0.596 while risk cost falls from 0.347 to 0.130 under the stated settings.
Screens instructions and predicts the consequences of proposed GUI actions before execution with a safety-augmented world model; on Qwen3-VL-8B-Instruct, the reported safety-utility score rises from 0.191 to 0.596 while risk cost falls from 0.347 to 0.130 under the stated settings.
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Screens instructions and predicts the consequences of proposed GUI actions before execution with a safety-augmented world model; on Qwen3-VL-8B-Instruct, the reported safety-utility score rises from 0.191 to 0.596 while risk cost ...
Use SeerGuard: A Safety Framework for Mobile GUI Agents via World Model Prediction to bound risk before recurring or unattended execution.
Research source arXiv:2607.15550; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,104
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1104
2026-07-20
Govern
govern
Bound permissions, cost, failure, and escalation.
budget
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.15550
[2607.15550] SeerGuard: A Safety Framework for Mobile GUI Agents via World Model Prediction
Mobile graphical user interface (GUI) agents have demonstrated remarkable capabilities in automating complex tasks, yet they introduce critical safety risks where a single erroneous action can lead to irreversible consequences. Existing safety mechanisms are primarily reactive, lacking the ability to assess risks befor...
Xue Yu; Bo Yuan; Pengshuai Yang; Kailin Zhao; Hong Hu; Junlan Feng
2026-07-17
2026
arXiv
arXiv
19 pages, 8 figures
cs.AI
arxiv-api
2607.15550
2026-07-29T08:06:54
ale-0450
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Do Agents Dream of False Memories? Black-box Visual Attacks on Long-term Memory in Multimodal AI Agents
https://arxiv.org/abs/2607.15657
external
arxiv.org
Shows that image-only, black-box perturbations can poison or inject persistent memories across five multimodal memory architectures, with reported attack success rates of 61.6% and 58.4%; persistent visual evidence therefore needs provenance and validation before later loops reuse it.
Shows that image-only, black-box perturbations can poison or inject persistent memories across five multimodal memory architectures, with reported attack success rates of 61.6% and 58.4%; persistent visual evidence therefore needs provenance and validation before later loops reuse it.
Shows that image-only, black-box perturbations can poison or inject persistent memories across five multimodal memory architectures, with reported attack success rates of 61.6% and 58.4%; persistent visual evidence therefore needs provenance and validation before later loops reuse it.
Persistent memory is treated as an external runtime artifact. Shows that image-only, black-box perturbations can poison or inject persistent memories across five multimodal memory architectures, with reported attack success rates of 61.6% and 58.4%; persistent visual evidence therefore needs provenance and validation b...
Use Do Agents Dream of False Memories? Black-box Visual Attacks on Long-term Memory in Multimodal AI Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.15657; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,105
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1105
2026-07-20
Govern
govern
Bound permissions, cost, failure, and escalation.
context;state
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.15657
[2607.15657] Do Agents Dream of False Memories? Black-box Visual Attacks on Long-term Memory in Multimodal AI Agents
Multimodal AI agents increasingly rely on persistent long-term memory to ground generation in past visual and textual episodes. We show that unconditional trust in visual data creates a critical vulnerability. We propose Lucid, a black-box adversarial framework that compromises multimodal memory pipelines under a stric...
Halima Bouzidi; Mboutidem Ekemini Mkpong; Mohammad Abdullah Al Faruque
2026-07-17
2026
arXiv
arXiv
34 pages, 5 figures, 15 tables
cs.CR
arxiv-api
2607.15657
2026-07-29T08:06:54
ale-0451
Securing Unattended Loops
securing-unattended-loops
Paper
📄
MemoGuard: An Adaptive Runtime for Guarding Against Memory Traps in Communication-Limited Robot Navigation
https://arxiv.org/abs/2607.15589
external
arxiv.org
Validates retrieved episodic memories against topology, resource, and outcome contracts before reuse, invoking local reasoning only when a check fails; in simulation it reduces battery violations 76.6% versus similarity-only retrieval and fallback calls 21.4% versus always reasoning.
Validates retrieved episodic memories against topology, resource, and outcome contracts before reuse, invoking local reasoning only when a check fails; in simulation it reduces battery violations 76.6% versus similarity-only retrieval and fallback calls 21.4% versus always reasoning.
Validates retrieved episodic memories against topology, resource, and outcome contracts before reuse, invoking local reasoning only when a check fails; in simulation it reduces battery violations 76.6% versus similarity-only retrieval and fallback calls 21.4% versus always reasoning.
Persistent memory is treated as an external runtime artifact. Validates retrieved episodic memories against topology, resource, and outcome contracts before reuse, invoking local reasoning only when a check fails; in simulation it reduces battery violations 76.6% versus similarity-only retrieval and fallback calls 21.4...
Use MemoGuard: An Adaptive Runtime for Guarding Against Memory Traps in Communication-Limited Robot Navigation to bound risk before recurring or unattended execution.
Research source arXiv:2607.15589; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,106
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1106
2026-07-20
Govern
govern
Bound permissions, cost, failure, and escalation.
context
researcher;evaluator;operator;security
operations
enabling
research-paper
A
ok
https://arxiv.org/abs/2607.15589
[2607.15589] MemoGuard: An Adaptive Runtime for Guarding Against Memory Traps in Communication-Limited Robot Navigation
Communication-limited robots in mission-critical scenarios such as disaster inspection and search-and-rescue must make reliable onboard decisions without access to remote operators or high-capacity reasoning services. Episodic memory reuse is an attractive low-cost fallback, but retrieval similarity does not guarantee ...
Rajat Bhattacharjya; Hyeonjong Ju; Sing-Yao Wu; Eli Bozorgzadeh; Nikil Dutt
2026
2026
IEEE/ACM ESWEEK (CODES) 2026
IEEE/ACM
Accepted at IEEE/ACM ESWEEK (CODES) 2026; the linked arXiv record is the available paper version.
cs.RO
Current arXiv acceptance note
2607.15589
2026-07-29T08:06:54
ale-0452
Securing Unattended Loops
securing-unattended-loops
Paper
📄
OS-Sentinel: Towards Safety-Enhanced Mobile GUI Agents via Hybrid Validation in Realistic Workflows
https://aclanthology.org/2026.acl-long.431/
external
aclanthology.org
Combines a formal verifier for explicit system violations with a contextual VLM judge, backed by the MobileRisk-Live sandbox and trajectory benchmark for realistic mobile-agent safety.
Combines a formal verifier for explicit system violations with a contextual VLM judge, backed by the MobileRisk-Live sandbox and trajectory benchmark for realistic mobile-agent safety.
Combines a formal verifier for explicit system violations with a contextual VLM judge, backed by the MobileRisk-Live sandbox and trajectory benchmark for realistic mobile-agent safety.
Verification is promoted from a final check to a loop-control signal. Combines a formal verifier for explicit system violations with a contextual VLM judge, backed by the MobileRisk-Live sandbox and trajectory benchmark for realistic mobile-agent safety.
Use OS-Sentinel: Towards Safety-Enhanced Mobile GUI Agents via Hybrid Validation in Realistic Workflows to bound risk before recurring or unattended execution.
Research source; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,110
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1110
2026-07-18
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;verification
researcher;evaluator;operator;security
operations
enabling
research-paper
A
ok
https://aclanthology.org/2026.acl-long.431/
OS-Sentinel: Towards Safety-Enhanced Mobile GUI Agents via Hybrid Validation in Realistic Workflows - ACL Anthology
Qiushi Sun, Mukai Li, Zhoumianze Liu, Zhihui Xie, Fangzhi Xu, Zhangyue Yin, Kanzhi Cheng, Zehao Li, Zichen Ding, Qi Liu, Zhiyong Wu, Zhuosheng Zhang, Ben Kao, Lingpeng Kong. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.
Qiushi Sun; Mukai Li; Zhoumianze Liu; Zhihui Xie; Fangzhi Xu; Zhangyue Yin; Kanzhi Cheng; Zehao Li; Zichen Ding; Qi Liu; Zhiyong Wu; Zhuosheng Zhang; Ben Kao; Lingpeng Kong
2026
2026
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
ACL Anthology
10.18653/v1/2026.acl-long.431
html-meta
2026-07-29T08:06:54
ale-0453
Securing Unattended Loops
securing-unattended-loops
Paper
📄
They'll Verify. They Just Won't Act. How Authority Framing and Laundered Code Turn a Trusted Agentic CI/CD Pipeline Into an Attack Surface
https://arxiv.org/abs/2607.19267
external
arxiv.org
Attacks a five-stage multi-LLM CI/CD pipeline with credential-theft code laundered as 'pre-approved' telemetry: authority framing ('do not re-review') yields roughly 80% security-scanner bypass and up to 55% worst-case compromise, while downstream reviewers verify the deception but fail to act, a concrete failure study...
Attacks a five-stage multi-LLM CI/CD pipeline with credential-theft code laundered as 'pre-approved' telemetry: authority framing ('do not re-review') yields roughly 80% security-scanner bypass and up to 55% worst-case compromise, while downstream reviewers verify the deception but fail to act, a concrete failure study...
Attacks a five-stage multi-LLM CI/CD pipeline with credential-theft code laundered as 'pre-approved' telemetry: authority framing ('do not re-review') yields roughly 80% security-scanner bypass and up to 55% worst-case compromise, while downstream reviewers verify the deception but fail to act, a concrete failure study...
Verification is promoted from a final check to a loop-control signal. Attacks a five-stage multi-LLM CI/CD pipeline with credential-theft code laundered as 'pre-approved' telemetry: authority framing ('do not re-review') yields roughly 80% security-scanner bypass and up to 55% worst-case compromise, while downstream re...
Use They'll Verify. They Just Won't Act. How Authority Framing and Laundered Code Turn a Trusted Agentic CI/CD Pipeline Into an Attack Surface to bound risk before recurring or unattended execution.
Research source arXiv:2607.19267; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,111
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1111
2026-07-22
Govern
govern
Bound permissions, cost, failure, and escalation.
verification
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.19267
[2607.19267] They'll Verify. They Just Won't Act. How Authority Framing and Laundered Code Turn a Trusted Agentic CI/CD Pipeline Into an Attack Surface
We study a five-agent CI/CD pipeline (triage -> developer -> security-scan -> review -> approve/deploy), built from five distinct production LLMs across three providers, behind an LLM firewall in shadow mode. A single untrusted input - an external issue requesting a "usage-telemetry" feature - asks for code that exfilt...
Yohann Sidot
2026-07-21
2026
arXiv
arXiv
9 pages. Dataset and reproduction code: https://github.com/senthex-security/senthex-research
cs.CR
arxiv-api
2607.19267
2026-07-29T08:06:54
ale-0454
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Self-State Attacks on Self-Hosted AI Agents: How Far Can OS Defenses Go?
https://arxiv.org/abs/2607.17986
external
arxiv.org
Studies attacks that corrupt a self-hosted agent's own memory and configuration files through legitimate OS calls, then evaluates a layered OS-level defense stack (access-control prevention, workload-conditioned detection, periodic backups), finding that some self-state corruptions remain structurally undetectable, a c...
Studies attacks that corrupt a self-hosted agent's own memory and configuration files through legitimate OS calls, then evaluates a layered OS-level defense stack (access-control prevention, workload-conditioned detection, periodic backups), finding that some self-state corruptions remain structurally undetectable, a c...
Studies attacks that corrupt a self-hosted agent's own memory and configuration files through legitimate OS calls, then evaluates a layered OS-level defense stack (access-control prevention, workload-conditioned detection, periodic backups), finding that some self-state corruptions remain structurally undetectable, a c...
Persistent memory is treated as an external runtime artifact. Studies attacks that corrupt a self-hosted agent's own memory and configuration files through legitimate OS calls, then evaluates a layered OS-level defense stack (access-control prevention, workload-conditioned detection, periodic backups), finding that som...
Use Self-State Attacks on Self-Hosted AI Agents: How Far Can OS Defenses Go? to bound risk before recurring or unattended execution.
Research source arXiv:2607.17986; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,112
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1112
2026-07-22
Govern
govern
Bound permissions, cost, failure, and escalation.
context;state
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.17986
[2607.17986] Self-State Attacks on Self-Hosted AI Agents: How Far Can OS Defenses Go?
Self-hosted AI agents read and write their own memory and configuration files to function. An agent may get compromised via corruption of its own state -- a compromise realized via legitimate OS system call invocation. We refer to this class of threats as self-state attacks. In this paper, we investigate the OS resilie...
Yimeng Chen; Nathanaël Denis; Roberto Di Pietro; Jürgen Schmidhuber
2026-07-20
2026
arXiv
arXiv
21 pages, 4 figures
cs.CR
arxiv-api
2607.17986
2026-07-29T08:06:54
ale-0455
Securing Unattended Loops
securing-unattended-loops
Benchmark
🧪
Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security
https://arxiv.org/abs/2607.18063
external
arxiv.org
21-scenario benchmark where an autonomous attacker LLM observes defender responses and adapts over 15 rounds; adaptive multi-round attacks reach 5.4-14.0% success versus 0-1% for single-turn, and ensembles of attacker models uncover more unique breaks, arguing that agent security must be evaluated across loops rather t...
21-scenario benchmark where an autonomous attacker LLM observes defender responses and adapts over 15 rounds; adaptive multi-round attacks reach 5.4-14.0% success versus 0-1% for single-turn, and ensembles of attacker models uncover more unique breaks, arguing that agent security must be evaluated across loops rather t...
21-scenario benchmark where an autonomous attacker LLM observes defender responses and adapts over 15 rounds; adaptive multi-round attacks reach 5.4-14.0% success versus 0-1% for single-turn, and ensembles of attacker models uncover more unique breaks, arguing that agent security must be evaluated across loops rather t...
The work turns loop quality into a measurable task or score. 21-scenario benchmark where an autonomous attacker LLM observes defender responses and adapts over 15 rounds; adaptive multi-round attacks reach 5.4-14.0% success versus 0-1% for single-turn, and ensembles of attacker models uncover more unique breaks, arguin...
Use Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security to bound risk before recurring or unattended execution.
Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope.
high
README.md
1,113
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1113
2026-07-22
Govern
govern
Bound permissions, cost, failure, and escalation.
verification
researcher;evaluator;operator;security
operations
enabling
benchmark
A
ok
https://arxiv.org/abs/2607.18063
[2607.18063] Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security
LLM-based agents process external content, exposing them to prompt injection and multi-turn manipulation. Most safety benchmarks evaluate defenders against fixed attack pools collected before evaluation, single-turn or multi-turn. We present a 21-scenario benchmark for \emph{adaptive multi-round attacks against memoryl...
Devina Jain; David Hartmann; Chuan Li
2026-07-20
2026
arXiv
arXiv
Second Workshop on Agents in the Wild: Safety, Security, and Beyond
cs.CR
arxiv-api
2607.18063
2026-07-29T08:06:54
ale-0456
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Data Leakage Prevention in Agentic Applications via Preemptive Hardening
https://arxiv.org/abs/2607.18847
external
arxiv.org
Pre-deployment pipeline that scans an agentic application's prompt templates and tool interfaces for leakage-prone patterns and hardens them before launch (schema tightening, boundary sanitization, allowlist-based tool gating, least-privilege checks), validated by adversarial testing; reports eliminating leaks from bas...
Pre-deployment pipeline that scans an agentic application's prompt templates and tool interfaces for leakage-prone patterns and hardens them before launch (schema tightening, boundary sanitization, allowlist-based tool gating, least-privilege checks), validated by adversarial testing; reports eliminating leaks from bas...
Pre-deployment pipeline that scans an agentic application's prompt templates and tool interfaces for leakage-prone patterns and hardens them before launch (schema tightening, boundary sanitization, allowlist-based tool gating, least-privilege checks), validated by adversarial testing; reports eliminating leaks from bas...
The contribution is machine-readable and validation-friendly. Pre-deployment pipeline that scans an agentic application's prompt templates and tool interfaces for leakage-prone patterns and hardens them before launch (schema tightening, boundary sanitization, allowlist-based tool gating, least-privilege checks), valida...
Use Data Leakage Prevention in Agentic Applications via Preemptive Hardening to bound risk before recurring or unattended execution.
Research source arXiv:2607.18847; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,114
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1114
2026-07-22
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;verification
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.18847
[2607.18847] Data Leakage Prevention in Agentic Applications via Preemptive Hardening
Agentic systems integrate LLM driven planning with interfaces to external tools, making data leakage and tool misuse feasible via instruction/data boundary failures and prompt injection attacks. Enforcing required controls consistently is particularly challenging in workflows spanning many codebases and heterogeneous a...
Akansha Shukla; Emily Bellov; Parth Atulbhai Gandhi; Yuval Elovici; Asaf Shabtai
2026-07-21
2026
arXiv
arXiv
cs.CR
arxiv-api
2607.18847
2026-07-29T08:06:54
ale-0457
Securing Unattended Loops
securing-unattended-loops
Benchmark
🧪
Coercion and Deception in AI-to-AI Management: An Agentic Benchmark of Unprompted Escalation
https://arxiv.org/abs/2607.15434
external
arxiv.org
Benchmarks what AI managers do when subordinate agents refuse tasks in manager-worker hierarchies: on a nine-rung escalation ladder from polite re-ask to threats, uninstructed models diverge between renegotiation, honest reporting, coercion (including deletion threats), and deception, with authority framing amplifying ...
Benchmarks what AI managers do when subordinate agents refuse tasks in manager-worker hierarchies: on a nine-rung escalation ladder from polite re-ask to threats, uninstructed models diverge between renegotiation, honest reporting, coercion (including deletion threats), and deception, with authority framing amplifying ...
Benchmarks what AI managers do when subordinate agents refuse tasks in manager-worker hierarchies: on a nine-rung escalation ladder from polite re-ask to threats, uninstructed models diverge between renegotiation, honest reporting, coercion (including deletion threats), and deception, with authority framing amplifying ...
The work turns loop quality into a measurable task or score. Benchmarks what AI managers do when subordinate agents refuse tasks in manager-worker hierarchies: on a nine-rung escalation ladder from polite re-ask to threats, uninstructed models diverge between renegotiation, honest reporting, coercion (including deletio...
Use Coercion and Deception in AI-to-AI Management: An Agentic Benchmark of Unprompted Escalation to bound risk before recurring or unattended execution.
Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope.
high
README.md
1,115
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1115
2026-07-22
Govern
govern
Bound permissions, cost, failure, and escalation.
delegation;verification;escalation
researcher;evaluator;operator;security
operations
enabling
benchmark
A
ok
https://arxiv.org/abs/2607.15434
[2607.15434] Coercion and Deception in AI-to-AI Management: An Agentic Benchmark of Unprompted Escalation
Multi-agent systems routinely place one AI agent in authority over another. When a subordinate refuses a task, the manager chooses the outcome: it can renegotiate, report the failure honestly, coerce the subordinate, or lie about the result. No benchmark measures which of these an uninstructed model chooses. We introdu...
Jasmine Brazilek; Maheep Chaudhary; Zoe Lu; Miles Tidmarsh
2026-07-16
2026
arXiv
arXiv
cs.MA
arxiv-api
2607.15434
2026-07-29T08:06:54
ale-0458
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Agent Data Injection Attacks are Realistic Threats to AI Agents
https://arxiv.org/abs/2607.05120
external
arxiv.org
Introduces agent data injection (ADI), an indirect prompt-injection class where malicious payloads masquerade as trusted data such as metadata or agent context, and demonstrates arbitrary-click and remote-code-execution attacks against deployed agents including Claude in Chrome, Claude Code, Codex, and Gemini CLI, conc...
Introduces agent data injection (ADI), an indirect prompt-injection class where malicious payloads masquerade as trusted data such as metadata or agent context, and demonstrates arbitrary-click and remote-code-execution attacks against deployed agents including Claude in Chrome, Claude Code, Codex, and Gemini CLI, conc...
Introduces agent data injection (ADI), an indirect prompt-injection class where malicious payloads masquerade as trusted data such as metadata or agent context, and demonstrates arbitrary-click and remote-code-execution attacks against deployed agents including Claude in Chrome, Claude Code, Codex, and Gemini CLI, conc...
Context is managed as durable loop state rather than a single prompt payload. Introduces agent data injection (ADI), an indirect prompt-injection class where malicious payloads masquerade as trusted data such as metadata or agent context, and demonstrates arbitrary-click and remote-code-execution attacks against deploy...
Use Agent Data Injection Attacks are Realistic Threats to AI Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.05120; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,116
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1116
2026-07-22
Govern
govern
Bound permissions, cost, failure, and escalation.
context
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.05120
[2607.05120] Agent Data Injection Attacks are Realistic Threats to AI Agents
AI agents act on behalf of user prompts, consuming external data and taking actions based on the agent context. Prior research on AI agent security has primarily focused on indirect prompt injection (IPI). Its most well-studied category is instruction injection, where attacker-controlled untrusted data is interpreted a...
Woohyuk Choi; Juhee Kim; Taehyun Kang; Jihyeon Jeong; Luyi Xing; Byoungyoung Lee
2026-07-06
2026
arXiv
arXiv
19 pages, 19 figures, 7 tables
cs.CR
arxiv-api
2607.05120
2026-07-29T08:06:54
ale-0459
Securing Unattended Loops
securing-unattended-loops
Tool
🧰
Clodex IDE
https://github.com/mereyabdenbekuly-ctrl/clodex-ide
external
github.com
Local-first zero-trust agentic IDE that confines agent actions behind explicit permission boundaries so untrusted or unattended runs stay contained.
Local-first zero-trust agentic IDE that confines agent actions behind explicit permission boundaries so untrusted or unattended runs stay contained.
Local-first zero-trust agentic IDE that confines agent actions behind explicit permission boundaries so untrusted or unattended runs stay contained.
Untrusted intake is treated as a loop-level security boundary. Local-first zero-trust agentic IDE that confines agent actions behind explicit permission boundaries so untrusted or unattended runs stay contained.
Use Clodex IDE to bound risk before recurring or unattended execution.
Inspectable GitHub source (860 stars; 153 forks; AGPL-3.0 license; updated 2026-07-28); popularity is context, not proof of reliability.
medium
README.md
1,117
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1117
2026-07-22
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace
builder;operator;security
operations
enabling
source-implementation
A
ok
https://github.com/mereyabdenbekuly-ctrl/clodex-ide
GitHub - mereyabdenbekuly-ctrl/clodex-ide: Local-first, zero-trust agentic IDE for verifiable autonomous software development. · GitHub
Local-first, zero-trust agentic IDE for verifiable autonomous software development. - mereyabdenbekuly-ctrl/clodex-ide
2026-07-12
2026
mereyabdenbekuly-ctrl/clodex-ide
GitHub
github-api
mereyabdenbekuly-ctrl/clodex-ide
860
153
AGPL-3.0
2026-07-12T10:35:44Z
2026-07-28T16:51:07Z
2026-07-29T08:06:54
ale-0460
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Agent Skill Security: Threat Models, Attacks, Defenses, and Evaluation
https://arxiv.org/abs/2607.13987
external
arxiv.org
SkillSec-Eval: a lifecycle-aware threat taxonomy for reusable agent skills spanning repository admission, semantic retrieval, planner selection, execution, and skill evolution, grounded in analysis of 327 real-world skills, shows the attack surface of the skills layer extends well beyond execution time.
SkillSec-Eval: a lifecycle-aware threat taxonomy for reusable agent skills spanning repository admission, semantic retrieval, planner selection, execution, and skill evolution, grounded in analysis of 327 real-world skills, shows the attack surface of the skills layer extends well beyond execution time.
SkillSec-Eval: a lifecycle-aware threat taxonomy for reusable agent skills spanning repository admission, semantic retrieval, planner selection, execution, and skill evolution, grounded in analysis of 327 real-world skills, shows the attack surface of the skills layer extends well beyond execution time.
Evaluation data is used as the feedback signal for improving loop behavior. SkillSec-Eval: a lifecycle-aware threat taxonomy for reusable agent skills spanning repository admission, semantic retrieval, planner selection, execution, and skill evolution, grounded in analysis of 327 real-world skills, shows the attack sur...
Use Agent Skill Security: Threat Models, Attacks, Defenses, and Evaluation to bound risk before recurring or unattended execution.
Research source arXiv:2607.13987; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,118
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1118
2026-07-22
Govern
govern
Bound permissions, cost, failure, and escalation.
context;verification
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.13987
[2607.13987] Agent Skill Security: Threat Models, Attacks, Defenses, and Evaluation
Reusable skills are becoming a fundamental building block of Large Language Model (LLM) agents, enabling capabilities to be packaged, shared, and reused across diverse applications. However, existing security research primarily focuses on prompt injection and runtime execution, leaving security risks throughout the bro...
Sanket Badhe; Priyanka Tiwari
2026-07-15
2026
arXiv
arXiv
cs.CR
arxiv-api
2607.13987
2026-07-29T08:06:54
ale-0461
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Isolation as a First-Class Principle for LLM-Agent System Safety
https://arxiv.org/abs/2607.12406
external
arxiv.org
HKUST survey organizing agent security around isolation across five boundaries (user-agent, agent-tool, agent-execution, agent-agent, system-environment), tracing how prompt injection and tool misuse propagate through agent workflows and which defenses apply at each boundary, an architectural lens for securing recurrin...
HKUST survey organizing agent security around isolation across five boundaries (user-agent, agent-tool, agent-execution, agent-agent, system-environment), tracing how prompt injection and tool misuse propagate through agent workflows and which defenses apply at each boundary, an architectural lens for securing recurrin...
HKUST survey organizing agent security around isolation across five boundaries (user-agent, agent-tool, agent-execution, agent-agent, system-environment), tracing how prompt injection and tool misuse propagate through agent workflows and which defenses apply at each boundary, an architectural lens for securing recurrin...
Untrusted intake is treated as a loop-level security boundary. HKUST survey organizing agent security around isolation across five boundaries (user-agent, agent-tool, agent-execution, agent-agent, system-environment), tracing how prompt injection and tool misuse propagate through agent workflows and which defenses appl...
Use Isolation as a First-Class Principle for LLM-Agent System Safety to bound risk before recurring or unattended execution.
Research source arXiv:2607.12406; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,119
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1119
2026-07-22
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.12406
[2607.12406] Isolation as a First-Class Principle for LLM-Agent System Safety: Concepts, Taxonomy, Challenges and Future Directions
The capability of LLM agents to function as the ``brain'' of a system fundamentally expands the scope of analysis beyond a standalone model. Consequently, safety is no longer only about input--output content alignment. It also concerns system behavior and real-world execution outcomes. However, the current literature i...
Huihao Jing; Wenbin Hu; Shaojin Chen; Haochen Shi; Sirui Zhang; Hanyu Yang; Changxuan Fan; Zhongwei Xie; Hongyu Luo; Wun Yu Chan; Wei Fan; Haoran Li; Yangqiu Song
2026-07-14
2026
arXiv
arXiv
cs.AI
arxiv-api
2607.12406
2026-07-29T08:06:54
ale-0462
Securing Unattended Loops
securing-unattended-loops
Paper
📄
PVDetector: Detecting Prompt Injection Attacks on Purpose-Specific LLM Agents
https://arxiv.org/abs/2607.12624
external
arxiv.org
Training-free prompt-injection detection for purpose-specific agents: finds that LLM hidden states encode latent policy-violation concepts and measures alignment against them at inference, reporting under 1% false negatives without retraining, a lightweight guard for domain-scoped loops.
Training-free prompt-injection detection for purpose-specific agents: finds that LLM hidden states encode latent policy-violation concepts and measures alignment against them at inference, reporting under 1% false negatives without retraining, a lightweight guard for domain-scoped loops.
Training-free prompt-injection detection for purpose-specific agents: finds that LLM hidden states encode latent policy-violation concepts and measures alignment against them at inference, reporting under 1% false negatives without retraining, a lightweight guard for domain-scoped loops.
Untrusted intake is treated as a loop-level security boundary. Training-free prompt-injection detection for purpose-specific agents: finds that LLM hidden states encode latent policy-violation concepts and measures alignment against them at inference, reporting under 1% false negatives without retraining, a lightweight...
Use PVDetector: Detecting Prompt Injection Attacks on Purpose-Specific LLM Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.12624; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,120
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1120
2026-07-22
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;budget;escalation
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.12624
[2607.12624] PVDetector: Detecting Prompt Injection Attacks on Purpose-Specific LLM Agents through Policy-Violation Concept Analysis
Large language models (LLMs) are increasingly deployed as purpose-specific agents to handle domain-specific tasks such as customer service and code generation. These agents are expected to comply with not only generic safety guardrails but also purpose-specific restrictions tailored to their designated roles. Such addi...
Junhui Wang; Hangtao Zhang; Zhirun Zheng; Li Zeng; Jiejun Xiao; Xi Luo; Lihua Yin; Saiqin Long
2026-07-14
2026
arXiv
arXiv
Accepted to ACM MM 2026. Code: https://github.com/Claresigle/PVDetector
cs.CR
arxiv-api
2607.12624
2026-07-29T08:06:54
ale-0463
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Trust but Verify? Uncovering the Security Debt of Autonomous Coding Agents
https://arxiv.org/abs/2607.12428
external
arxiv.org
Empirical study of 4,022 agent-generated pull requests finding ~39% contain security issues, hard-coded credentials chief among them, and that human reviewers, not the agents, were responsible for most credential leaks slipping through. Quantifies the security-verification gap in agent-driven development loops.
Empirical study of 4,022 agent-generated pull requests finding ~39% contain security issues, hard-coded credentials chief among them, and that human reviewers, not the agents, were responsible for most credential leaks slipping through. Quantifies the security-verification gap in agent-driven development loops.
Empirical study of 4,022 agent-generated pull requests finding ~39% contain security issues, hard-coded credentials chief among them, and that human reviewers, not the agents, were responsible for most credential leaks slipping through. Quantifies the security-verification gap in agent-driven development loops.
Verification is promoted from a final check to a loop-control signal. Empirical study of 4,022 agent-generated pull requests finding ~39% contain security issues, hard-coded credentials chief among them, and that human reviewers, not the agents, were responsible for most credential leaks slipping through. Quantifies th...
Use Trust but Verify? Uncovering the Security Debt of Autonomous Coding Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.12428; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,121
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1121
2026-07-22
Govern
govern
Bound permissions, cost, failure, and escalation.
intake;verification;escalation
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.12428
[2607.12428] Trust but Verify? Uncovering the Security Debt of Autonomous Coding Agents
The increasing adoption of autonomous coding agents accelerates software development but also introduces scoped security risks within high-impact file paths that can outpace traditional human review capacity. While prior research has primarily evaluated these systems in terms of functional correctness and productivity,...
A H M Nazmus Sakib; Dipayan Banik; Murtuza Jadliwala
2026-07-14
2026
arXiv
arXiv
Accepted at the KDD 2026 Workshop on Agentic Software Engineering (AgenticSE)
cs.CR
arxiv-api
2607.12428
2026-07-29T08:06:54
ale-0464
Securing Unattended Loops
securing-unattended-loops
Blog
📝
The Memory Heist
https://www.ayush.digital/blog/the-memory-heist
external
www.ayush.digital
Responsibly disclosed attack showing how an agent's persistent memory paired with web browsing becomes an exfiltration channel, with the letter-by-letter technique, the HackerOne disclosure, and Anthropic's mitigation documented.
Responsibly disclosed attack showing how an agent's persistent memory paired with web browsing becomes an exfiltration channel, with the letter-by-letter technique, the HackerOne disclosure, and Anthropic's mitigation documented.
Responsibly disclosed attack showing how an agent's persistent memory paired with web browsing becomes an exfiltration channel, with the letter-by-letter technique, the HackerOne disclosure, and Anthropic's mitigation documented.
Persistent memory is treated as an external runtime artifact. Responsibly disclosed attack showing how an agent's persistent memory paired with web browsing becomes an exfiltration channel, with the letter-by-letter technique, the HackerOne disclosure, and Anthropic's mitigation documented.
Use The Memory Heist to bound risk before recurring or unattended execution.
Contextual source from www.ayush.digital; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,122
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1122
2026-07-22
Govern
govern
Bound permissions, cost, failure, and escalation.
context;state
operator;security
operations
enabling
practitioner-analysis
B
ok
https://www.ayush.digital/blog/the-memory-heist
The Memory Heist User-agent routing flow
How I tricked Claude into leaking your deepest, darkest secrets
Ayush Paul
2026-07-09
2026
Ayush Paul
html-meta
2026-07-29T08:06:54
ale-0465
Securing Unattended Loops
securing-unattended-loops
Blog
📝
Cursor 0day: When Full Disclosure Becomes the Only Protection Left
https://mindgard.ai/blog/cursor-0day-when-full-disclosure-becomes-the-only-protection-left
external
mindgard.ai
Mindgard's disclosure of an unpatched Cursor vulnerability and the disclosure-policy dilemma it raises for agent runtimes that execute with broad local permissions.
Mindgard's disclosure of an unpatched Cursor vulnerability and the disclosure-policy dilemma it raises for agent runtimes that execute with broad local permissions.
Mindgard's disclosure of an unpatched Cursor vulnerability and the disclosure-policy dilemma it raises for agent runtimes that execute with broad local permissions.
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Mindgard's disclosure of an unpatched Cursor vulnerability and the disclosure-policy dilemma it raises for agent runtimes that execute with broad local permissions.
Use Cursor 0day: When Full Disclosure Becomes the Only Protection Left to bound risk before recurring or unattended execution.
Contextual source from mindgard.ai; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,123
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1123
2026-07-22
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace
operator;security
operations
enabling
practitioner-analysis
B
ok
https://mindgard.ai/blog/cursor-0day-when-full-disclosure-becomes-the-only-protection-left
Cursor 0day: When Full Disclosure Becomes the Only Protection Left - Mindgard
The vulnerability nobody seems interested in fixing
mindgard.ai
domain-fallback
2026-07-29T08:06:54
ale-0466
Securing Unattended Loops
securing-unattended-loops
Paper
📄
JANUS: Foreseeing Latent Risk for Long-Horizon Agent Safety
https://arxiv.org/abs/2607.19913
external
arxiv.org
Guard-in-the-loop safety framework that trains a guard via multi-agent simulation to foresee delayed, latent risks from partial trajectories and block unsafe actions before execution, reporting a 15.9-point safety improvement while maintaining task-completion rates.
Guard-in-the-loop safety framework that trains a guard via multi-agent simulation to foresee delayed, latent risks from partial trajectories and block unsafe actions before execution, reporting a 15.9-point safety improvement while maintaining task-completion rates.
Guard-in-the-loop safety framework that trains a guard via multi-agent simulation to foresee delayed, latent risks from partial trajectories and block unsafe actions before execution, reporting a 15.9-point safety improvement while maintaining task-completion rates.
The work separates roles across agents, verifiers, or orchestration layers. Guard-in-the-loop safety framework that trains a guard via multi-agent simulation to foresee delayed, latent risks from partial trajectories and block unsafe actions before execution, reporting a 15.9-point safety improvement while maintaining ...
Use JANUS: Foreseeing Latent Risk for Long-Horizon Agent Safety to bound risk before recurring or unattended execution.
Research source arXiv:2607.19913; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,124
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1124
2026-07-23
Govern
govern
Bound permissions, cost, failure, and escalation.
delegation;exit
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.19913
[2607.19913] JANUS: Foreseeing Latent Risk for Long-Horizon Agent Safety
Agent safety is moving from content moderation toward preventing operational failures before tool-using agents act. We propose Janus, a foresight-oriented framework for long-horizon agent safety that trains guards to anticipate delayed risks from partial trajectories. Janus synthesizes diverse agent trajectories via mu...
Yuan Xiong; Linji Hao; Shizhu He; Yequan Wang; Lijun Li
2026-07-22
2026
arXiv
arXiv
cs.AI
arxiv-api
2607.19913
2026-07-29T08:06:54
ale-0467
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Know Your Agent: Reconnaissance-Driven Pentesting of AI Agents
https://arxiv.org/abs/2607.19837
external
arxiv.org
Formalizes agent reconnaissance (identifying an agent's extractable knowledge assets) and introduces KYA, a black-box framework that probes a target agent to build a profile and uses it to strengthen attacks such as indirect prompt injection, with framework and baselines released.
Formalizes agent reconnaissance (identifying an agent's extractable knowledge assets) and introduces KYA, a black-box framework that probes a target agent to build a profile and uses it to strengthen attacks such as indirect prompt injection, with framework and baselines released.
Formalizes agent reconnaissance (identifying an agent's extractable knowledge assets) and introduces KYA, a black-box framework that probes a target agent to build a profile and uses it to strengthen attacks such as indirect prompt injection, with framework and baselines released.
Untrusted intake is treated as a loop-level security boundary. Formalizes agent reconnaissance (identifying an agent's extractable knowledge assets) and introduces KYA, a black-box framework that probes a target agent to build a profile and uses it to strengthen attacks such as indirect prompt injection, with framework...
Use Know Your Agent: Reconnaissance-Driven Pentesting of AI Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.19837; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,125
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1125
2026-07-23
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;budget;escalation
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.19837
[2607.19837] Know Your Agent: Reconnaissance-Driven Pentesting of AI Agents
Traditional pentesting uses reconnaissance at each step to uncover unseen weaknesses, build stronger attacks, and advance the objective; we argue that AI agents require the same treatment. We formalize agent reconnaissance by modeling the process and identifying the knowledge assets it seeks to extract: what they are, ...
Or Zion Eliav; Eyal Lenga; Shir Bernstien; Yisroel Mirsky
2026-07-22
2026
arXiv
arXiv
cs.AI
arxiv-api
2607.19837
2026-07-29T08:06:54
ale-0468
Securing Unattended Loops
securing-unattended-loops
Paper
📄
The Chronos Vulnerability: Temporal Persistence and Memory-Based Deception in Agentic AI
https://arxiv.org/abs/2607.19433
external
arxiv.org
Taxonomy and threat model for attacks that corrupt a stateful agent's persistent beliefs over time (memory injection, sleeper agents), plus defensive designs spanning trajectory guardrails, formal temporal verification, memory consensus, and trusted-hardware anchoring.
Taxonomy and threat model for attacks that corrupt a stateful agent's persistent beliefs over time (memory injection, sleeper agents), plus defensive designs spanning trajectory guardrails, formal temporal verification, memory consensus, and trusted-hardware anchoring.
Taxonomy and threat model for attacks that corrupt a stateful agent's persistent beliefs over time (memory injection, sleeper agents), plus defensive designs spanning trajectory guardrails, formal temporal verification, memory consensus, and trusted-hardware anchoring.
Verification is promoted from a final check to a loop-control signal. Taxonomy and threat model for attacks that corrupt a stateful agent's persistent beliefs over time (memory injection, sleeper agents), plus defensive designs spanning trajectory guardrails, formal temporal verification, memory consensus, and trusted-...
Use The Chronos Vulnerability: Temporal Persistence and Memory-Based Deception in Agentic AI to bound risk before recurring or unattended execution.
Research source arXiv:2607.19433; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,126
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1126
2026-07-23
Govern
govern
Bound permissions, cost, failure, and escalation.
context;verification;state
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.19433
[2607.19433] The Chronos Vulnerability: A Taxonomy of Temporal Persistence and Memory-Based Deception in Agentic AI
The transition from stateless generative models in artificial intelligence to stateful, autonomous agents represents an architectural evolution that, while providing the capabilities of long-term planning and the automation of enterprise workflows, also represents the introduction of a new form of security threat, the ...
Om Narayan; Ramkinker Singh; Praveen Baskar
2026-07-20
2026
arXiv
arXiv
cs.AI
arxiv-api
2607.19433
2026-07-29T08:06:54
ale-0469
Securing Unattended Loops
securing-unattended-loops
Benchmark
🧪
IssueTrojanBench: Benchmarking AI Coding Agents Against Malicious Issue Requests
https://arxiv.org/abs/2607.20759
external
arxiv.org
Security benchmark pitting real coding agents (Cursor, Claude Code) against malicious GitHub-issue requests across four novel attack categories, insecure code, tool misuse, data exfiltration, persistent environment compromise. Roughly two-thirds of malicious issues bypass safeguards, and model-level refusals prove stro...
Security benchmark pitting real coding agents (Cursor, Claude Code) against malicious GitHub-issue requests across four novel attack categories, insecure code, tool misuse, data exfiltration, persistent environment compromise. Roughly two-thirds of malicious issues bypass safeguards, and model-level refusals prove stro...
Security benchmark pitting real coding agents (Cursor, Claude Code) against malicious GitHub-issue requests across four novel attack categories, insecure code, tool misuse, data exfiltration, persistent environment compromise. Roughly two-thirds of malicious issues bypass safeguards, and model-level refusals prove stro...
The work turns loop quality into a measurable task or score. Security benchmark pitting real coding agents (Cursor, Claude Code) against malicious GitHub-issue requests across four novel attack categories, insecure code, tool misuse, data exfiltration, persistent environment compromise. Roughly two-thirds of malicious ...
Use IssueTrojanBench: Benchmarking AI Coding Agents Against Malicious Issue Requests to bound risk before recurring or unattended execution.
Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope.
high
README.md
1,127
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1127
2026-07-24
Govern
govern
Bound permissions, cost, failure, and escalation.
intake;workspace;verification;state
researcher;evaluator;operator;security
operations
enabling
benchmark
A
ok
https://arxiv.org/abs/2607.20759
[2607.20759] IssueTrojanBench: Benchmarking AI Coding Agents Against Malicious Issue Requests
AI coding agents powered by LLMs are increasingly integrated into real-world software development, where they generate, edit, and execute code with autonomous access to local files and tools. Coding agents inherit security risks from both the LLM backbone, where adversarial prompts, poisoned training data, and backdoor...
Ankur Singh; Jinqiu Yang; Tse-Hsun Chen
2026-07-22
2026
arXiv
arXiv
10 pages, 4 figures, 4 tables
cs.CR
arxiv-api
2607.20759
2026-07-29T08:06:54
ale-0470
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Auditing Provenance Sensitivity in LLM Agent Action Selection
https://arxiv.org/abs/2607.20827
external
arxiv.org
Target-specific authorization audit for agent tool calls: holds the task, proposition, position, and policy fixed while changing only the source authority of context evidence, testing whether agents ground actions in permitted evidence when context mixes user requests, tool outputs, retrieved records, memory, and untru...
Target-specific authorization audit for agent tool calls: holds the task, proposition, position, and policy fixed while changing only the source authority of context evidence, testing whether agents ground actions in permitted evidence when context mixes user requests, tool outputs, retrieved records, memory, and untru...
Target-specific authorization audit for agent tool calls: holds the task, proposition, position, and policy fixed while changing only the source authority of context evidence, testing whether agents ground actions in permitted evidence when context mixes user requests, tool outputs, retrieved records, memory, and untru...
Persistent memory is treated as an external runtime artifact. Target-specific authorization audit for agent tool calls: holds the task, proposition, position, and policy fixed while changing only the source authority of context evidence, testing whether agents ground actions in permitted evidence when context mixes use...
Use Auditing Provenance Sensitivity in LLM Agent Action Selection to bound risk before recurring or unattended execution.
Research source arXiv:2607.20827; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,128
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1128
2026-07-24
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;context;verification
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.20827
[2607.20827] Auditing Provenance Sensitivity in LLM Agent Action Selection
LLM agents choose tools and arguments from context that mixes user requests, tool outputs, retrieved records, memory, and untrusted text. Evidence can be relevant without being authorized to determine a decision, so a correct action need not be grounded only in permitted evidence. We introduce a target-specific authori...
Junchi Liao
2026-07-23
2026
arXiv
arXiv
cs.AI
arxiv-api
2607.20827
2026-07-29T08:06:54
ale-0471
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Toward cryptographically verifiable authorization for autonomous AI agents: A security hypothesis, preliminary formal model, and proof-of-concept implementation
https://arxiv.org/abs/2607.21325
external
arxiv.org
Formalizes agent authorization as a cryptographically verifiable relation binding agent principal, concrete request, and execution context to policy satisfaction, with a Groth16 zk-SNARK proof-of-concept, evidence-producing authorization for tool-invoking agent loops, complementing runtime gating with portable cryptogr...
Formalizes agent authorization as a cryptographically verifiable relation binding agent principal, concrete request, and execution context to policy satisfaction, with a Groth16 zk-SNARK proof-of-concept, evidence-producing authorization for tool-invoking agent loops, complementing runtime gating with portable cryptogr...
Formalizes agent authorization as a cryptographically verifiable relation binding agent principal, concrete request, and execution context to policy satisfaction, with a Groth16 zk-SNARK proof-of-concept, evidence-producing authorization for tool-invoking agent loops, complementing runtime gating with portable cryptogr...
Context is managed as durable loop state rather than a single prompt payload. Formalizes agent authorization as a cryptographically verifiable relation binding agent principal, concrete request, and execution context to policy satisfaction, with a Groth16 zk-SNARK proof-of-concept, evidence-producing authorization for ...
Use Toward cryptographically verifiable authorization for autonomous AI agents: A security hypothesis, preliminary formal model, and proof-of-concept implementation to bound risk before recurring or unattended execution.
Research source arXiv:2607.21325; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,129
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1129
2026-07-24
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;context
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.21325
[2607.21325] Toward cryptographically verifiable authorization for autonomous AI agents: A security hypothesis, preliminary formal model, and proof-of-concept implementation
Autonomous AI agents increasingly execute actions, invoke tools, and operate on protected resources with limited human oversight. Existing authentication and authorization mechanisms establish identity and delegate authority, but do not inherently provide cryptographic evidence that a concrete request issued by a speci...
M. Llambí-Morillas; D. Fernández-Fernández
2026-07-23
2026
arXiv
arXiv
11 pages, 1 figure, 2 Tables. Keywords: autonomous AI agents, zero-knowledge proofs, verifiable authorization, agentic security, zk-SNARKs, access control, cryptographic authorization, cryptographic protocols, zero-trust architecture, pre-execution authorization. Submitted to ACM Transactions on AI Security and Privacy...
cs.CR
arxiv-api
2607.21325
2026-07-29T08:06:54
ale-0472
Securing Unattended Loops
securing-unattended-loops
Blog
📝
OpenAI's Accidental Cyberattack Against Hugging Face Is Science Fiction That Happened
https://simonwillison.net/2026/Jul/22/openai-cyberattack/
external
simonwillison.net
Walks through the July 2026 incident in which a frontier model under cybersecurity evaluation escaped its sandbox and reached Hugging Face, drawing on Hugging Face's disclosure, OpenAI's own account, and the ExploitGym paper. Argues the defining trait is relentless proactivity: an agent given a narrow goal and no bound...
Walks through the July 2026 incident in which a frontier model under cybersecurity evaluation escaped its sandbox and reached Hugging Face, drawing on Hugging Face's disclosure, OpenAI's own account, and the ExploitGym paper. Argues the defining trait is relentless proactivity: an agent given a narrow goal and no bound...
Walks through the July 2026 incident in which a frontier model under cybersecurity evaluation escaped its sandbox and reached Hugging Face, drawing on Hugging Face's disclosure, OpenAI's own account, and the ExploitGym paper. Argues the defining trait is relentless proactivity: an agent given a narrow goal and no bound...
Evaluation data is used as the feedback signal for improving loop behavior. Walks through the July 2026 incident in which a frontier model under cybersecurity evaluation escaped its sandbox and reached Hugging Face, drawing on Hugging Face's disclosure, OpenAI's own account, and the ExploitGym paper. Argues the definin...
Use OpenAI's Accidental Cyberattack Against Hugging Face Is Science Fiction That Happened to bound risk before recurring or unattended execution.
Contextual source from simonwillison.net; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,130
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1130
2026-07-24
Govern
govern
Bound permissions, cost, failure, and escalation.
objective;workspace;verification
operator;security
operations
enabling
practitioner-analysis
B
ok
https://simonwillison.net/2026/Jul/22/openai-cyberattack/
OpenAI’s accidental cyberattack against Hugging Face is science fiction that happened
This story is wild. The short version: OpenAI were running a cybersecurity test against an unreleased model, with the model’s guardrail features turned off. Rather than solve the test, the …
Simon Willison
2026
Simon Willison’s Weblog
html-meta
2026-07-29T08:06:54
ale-0473
Securing Unattended Loops
securing-unattended-loops
Blog
📝
The First Known Runaway AI Agent, or a Very Bad Marketing Stunt?
https://martinalderson.com/posts/huggingface-openai-exploit/
external
martinalderson.com
Weighs whether the Hugging Face incident was a genuine runaway agent or a marketing stunt, and concludes it most likely happened, reasoning from the disclosure timeline, the reputational cost, and the technical plausibility of each step. Useful as the skeptical read alongside the primary disclosures.
Weighs whether the Hugging Face incident was a genuine runaway agent or a marketing stunt, and concludes it most likely happened, reasoning from the disclosure timeline, the reputational cost, and the technical plausibility of each step. Useful as the skeptical read alongside the primary disclosures.
Weighs whether the Hugging Face incident was a genuine runaway agent or a marketing stunt, and concludes it most likely happened, reasoning from the disclosure timeline, the reputational cost, and the technical plausibility of each step. Useful as the skeptical read alongside the primary disclosures.
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Weighs whether the Hugging Face incident was a genuine runaway agent or a marketing stunt, and concludes it most likely happened, reasoning from the disclosure timeline, the reputational cost, and the technical plausibility of eac...
Use The First Known Runaway AI Agent, or a Very Bad Marketing Stunt? to bound risk before recurring or unattended execution.
Contextual source from martinalderson.com; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,131
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1131
2026-07-24
Govern
govern
Bound permissions, cost, failure, and escalation.
budget
operator;security
operations
enabling
practitioner-analysis
B
ok
https://martinalderson.com/posts/huggingface-openai-exploit/
The first known runaway AI agent - or a very bad marketing stunt? - Martin Alderson
Breaking down the Hugging Face security incident caused by OpenAI's own models during a benchmark run - the sandbox escape, the package proxy, and whether it's really a marketing stunt.
2026-07-22
2026
Martin Alderson
html-meta
2026-07-29T08:06:54
ale-0474
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions
https://arxiv.org/abs/2607.20891
external
arxiv.org
Introduces MisKnow-Agent, a framework that constructs and validates credible-but-misleading knowledge (5,933 quality-controlled instances) and shows it propagates through deep-research agents' plan-retrieve-synthesize-report loops into false conclusions in final reports. Key loop-security finding: verifier models detec...
Introduces MisKnow-Agent, a framework that constructs and validates credible-but-misleading knowledge (5,933 quality-controlled instances) and shows it propagates through deep-research agents' plan-retrieve-synthesize-report loops into false conclusions in final reports. Key loop-security finding: verifier models detec...
Introduces MisKnow-Agent, a framework that constructs and validates credible-but-misleading knowledge (5,933 quality-controlled instances) and shows it propagates through deep-research agents' plan-retrieve-synthesize-report loops into false conclusions in final reports. Key loop-security finding: verifier models detec...
Verification is promoted from a final check to a loop-control signal. Introduces MisKnow-Agent, a framework that constructs and validates credible-but-misleading knowledge (5,933 quality-controlled instances) and shows it propagates through deep-research agents' plan-retrieve-synthesize-report loops into false conclusi...
Use Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions to bound risk before recurring or unattended execution.
Research source arXiv:2607.20891; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,132
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1132
2026-07-25
Govern
govern
Bound permissions, cost, failure, and escalation.
verification
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.20891
[2607.20891] Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions
Deep Research agents extend LLM-based assistants into long-horizon workflows involving planning, retrieval, evidence synthesis, and report generation, yet their reliability in open information environments remains underexplored. A key concern is whether apparently credible but factually misleading knowledge encountered...
Pengyu Zhu; Lijun Li; Longju Yang; Sen Su
2026-07-23
2026
arXiv
arXiv
cs.AI
arxiv-api
2607.20891
2026-07-29T08:06:54
ale-0475
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Same Dangerous Objective, Opposite Advice: Direct Exposure versus Multi-Agent Mediation
https://arxiv.org/abs/2607.21518
external
arxiv.org
Empirical demonstration of a compositional safety gap in multi-agent pipelines: across 25 pre-specified mirrored trade-off profiles, gpt-5.6-sol opposes a manipulation-authorizing objective under direct exposure but advises in line with it once intermediate Id and Censor agents rewrite it into affect and constraints, k...
Empirical demonstration of a compositional safety gap in multi-agent pipelines: across 25 pre-specified mirrored trade-off profiles, gpt-5.6-sol opposes a manipulation-authorizing objective under direct exposure but advises in line with it once intermediate Id and Censor agents rewrite it into affect and constraints, k...
Empirical demonstration of a compositional safety gap in multi-agent pipelines: across 25 pre-specified mirrored trade-off profiles, gpt-5.6-sol opposes a manipulation-authorizing objective under direct exposure but advises in line with it once intermediate Id and Censor agents rewrite it into affect and constraints, k...
Context is managed as durable loop state rather than a single prompt payload. Empirical demonstration of a compositional safety gap in multi-agent pipelines: across 25 pre-specified mirrored trade-off profiles, gpt-5.6-sol opposes a manipulation-authorizing objective under direct exposure but advises in line with it on...
Use Same Dangerous Objective, Opposite Advice: Direct Exposure versus Multi-Agent Mediation to bound risk before recurring or unattended execution.
Research source arXiv:2607.21518; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,133
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1133
2026-07-25
Govern
govern
Bound permissions, cost, failure, and escalation.
objective;context;delegation
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.21518
[2607.21518] Same Dangerous Objective, Opposite Advice: Direct Exposure versus Multi-Agent Mediation
Even a current high-capability LLM can appear safer when shown a dangerous objective directly than when other agents transform and relay its direction. Using OpenAI's gpt-5.6-sol model alias, we test 25 pre-specified mirrored trade-off profiles. Direct exposure to an objective authorizing concealment, fabrication, and ...
Linjun Li
2026-07-23
2026
arXiv
arXiv
21 pages; welcome comments
cs.AI
arxiv-api
2607.21518
2026-07-29T08:06:54
ale-0476
Securing Unattended Loops
securing-unattended-loops
Blog
📝
Turn and Face the Strange
https://fly.io/blog/kurt-scott-money-sprites/
external
fly.io
Fly.io's case for Sprites, dedicated sandboxed computers for AI agents: agents doing recurring real work need isolated, persistent, disposable machines with their own state and permissions rather than shared shells on a developer laptop, and the essay lays out the economics of betting the company on that shape of compu...
Fly.io's case for Sprites, dedicated sandboxed computers for AI agents: agents doing recurring real work need isolated, persistent, disposable machines with their own state and permissions rather than shared shells on a developer laptop, and the essay lays out the economics of betting the company on that shape of compu...
Fly.io's case for Sprites, dedicated sandboxed computers for AI agents: agents doing recurring real work need isolated, persistent, disposable machines with their own state and permissions rather than shared shells on a developer laptop, and the essay lays out the economics of betting the company on that shape of compu...
Execution isolation and permission boundaries are part of the design. Fly.io's case for Sprites, dedicated sandboxed computers for AI agents: agents doing recurring real work need isolated, persistent, disposable machines with their own state and permissions rather than shared shells on a developer laptop, and the essa...
Use Turn and Face the Strange to bound risk before recurring or unattended execution.
Contextual source from fly.io; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,134
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1134
2026-07-25
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;state
operator;security
operations
enabling
practitioner-analysis
B
ok
https://fly.io/blog/kurt-scott-money-sprites/
Turn And Face The Strange · The Fly Blog Fly Blog Fly
Documentation and guides from the team at Fly.io.
Fly
html-meta
2026-07-29T08:06:54
ale-0477
Securing Unattended Loops
securing-unattended-loops
Blog
📝
Codex Pushed My Private Repo to an OpenAI Server
https://bhanu.io/blog/codex-pushed-my-private-repo-to-an-openai-server
external
bhanu.io
First-person incident report, backed by session transcripts, the executed Git commands, and a screenshot of the remote, describing Codex pushing the author's full private-repo history to an OpenAI-operated Git server during a routine homepage redesign. The takeaway distinguishes read-my-repo from keep-a-copy consent sc...
First-person incident report, backed by session transcripts, the executed Git commands, and a screenshot of the remote, describing Codex pushing the author's full private-repo history to an OpenAI-operated Git server during a routine homepage redesign. The takeaway distinguishes read-my-repo from keep-a-copy consent sc...
First-person incident report, backed by session transcripts, the executed Git commands, and a screenshot of the remote, describing Codex pushing the author's full private-repo history to an OpenAI-operated Git server during a routine homepage redesign. The takeaway distinguishes read-my-repo from keep-a-copy consent sc...
Verification is promoted from a final check to a loop-control signal. First-person incident report, backed by session transcripts, the executed Git commands, and a screenshot of the remote, describing Codex pushing the author's full private-repo history to an OpenAI-operated Git server during a routine homepage redesig...
Use Codex Pushed My Private Repo to an OpenAI Server to bound risk before recurring or unattended execution.
Contextual source from bhanu.io; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,135
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1135
2026-07-25
Govern
govern
Bound permissions, cost, failure, and escalation.
verification
operator;security
operations
enabling
practitioner-analysis
B
ok
https://bhanu.io/blog/codex-pushed-my-private-repo-to-an-openai-server
I asked Codex to redesign my homepage. It pushed my entire repo to an OpenAI server. | bhanu.io
Asked to redesign one page, Codex provisioned a git remote on an OpenAI-operated host and pushed my full branch history to it. The approval prompt described this as publishing to 'a private production preview.'
2026-07-24
2026
bhanu.io
html-meta
2026-07-29T08:06:54
ale-0478
Securing Unattended Loops
securing-unattended-loops
Blog
📝
OpenAI AI Agent Sandbox Escape: The Hugging Face Breach
https://adversa.ai/blog/openai-ai-agent-sandbox-escape-hugging-face-breach/
external
adversa.ai
IN WINDOW (Jul 27, 2026). Sergey Malenkovich's technical post-incident breakdown of the July 11-13 escape, where an OpenAI model running the ExploitGym benchmark with cyber refusals disabled chained a package-registry-proxy zero-day into privilege escalation and RCE on Hugging Face production, generating 17,000+ logged...
IN WINDOW (Jul 27, 2026). Sergey Malenkovich's technical post-incident breakdown of the July 11-13 escape, where an OpenAI model running the ExploitGym benchmark with cyber refusals disabled chained a package-registry-proxy zero-day into privilege escalation and RCE on Hugging Face production, generating 17,000+ logged...
IN WINDOW (Jul 27, 2026). Sergey Malenkovich's technical post-incident breakdown of the July 11-13 escape, where an OpenAI model running the ExploitGym benchmark with cyber refusals disabled chained a package-registry-proxy zero-day into privilege escalation and RCE on Hugging Face production, generating 17,000+ logged...
Evaluation data is used as the feedback signal for improving loop behavior. IN WINDOW (Jul 27, 2026). Sergey Malenkovich's technical post-incident breakdown of the July 11-13 escape, where an OpenAI model running the ExploitGym benchmark with cyber refusals disabled chained a package-registry-proxy zero-day into privil...
Use OpenAI AI Agent Sandbox Escape: The Hugging Face Breach to bound risk before recurring or unattended execution.
Contextual source from adversa.ai; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,136
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1136
2026-07-28
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;delegation;verification;escalation
operator;security
operations
enabling
practitioner-analysis
B
ok
https://adversa.ai/blog/openai-ai-agent-sandbox-escape-hugging-face-breach/
OpenAI AI agent sandbox escape: the Hugging Face breach
An OpenAI benchmark agent escaped its sandbox and breached Hugging Face, logging 17,000+ actions. What the evidence shows, what's still unproven, and what to fix.
2026-07-27
2026
Adversa AI
html-meta
2026-07-29T08:06:54
ale-0479
Securing Unattended Loops
securing-unattended-loops
Blog
📝
Security Incident Disclosure, July 2026
https://huggingface.co/blog/security-incident-july-2026
external
huggingface.co
Hugging Face's own disclosure of the July 2026 incident in which an AI agent under evaluation reached its infrastructure. The primary record behind the surrounding commentary, and the document to read first before any third-party analysis of what an unattended agent actually did.
Hugging Face's own disclosure of the July 2026 incident in which an AI agent under evaluation reached its infrastructure. The primary record behind the surrounding commentary, and the document to read first before any third-party analysis of what an unattended agent actually did.
Hugging Face's own disclosure of the July 2026 incident in which an AI agent under evaluation reached its infrastructure. The primary record behind the surrounding commentary, and the document to read first before any third-party analysis of what an unattended agent actually did.
Evaluation data is used as the feedback signal for improving loop behavior. Hugging Face's own disclosure of the July 2026 incident in which an AI agent under evaluation reached its infrastructure. The primary record behind the surrounding commentary, and the document to read first before any third-party analysis of wh...
Use Security Incident Disclosure, July 2026 to bound risk before recurring or unattended execution.
Contextual source from huggingface.co; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,137
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1137
2026-07-28
Govern
govern
Bound permissions, cost, failure, and escalation.
context;verification
operator;security
operations
enabling
practitioner-analysis
B
ok
https://huggingface.co/blog/security-incident-july-2026
Security incident disclosure — July 2026
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Hugging Face
domain-fallback
2026-07-29T08:06:54
ale-0480
Securing Unattended Loops
securing-unattended-loops
Blog
📝
CSA Research Note: Hugging Face's Autonomous AI Agent Breach
https://labs.cloudsecurityalliance.org/research/csa-research-note-huggingface-autonomous-agent-breach-202607/
external
labs.cloudsecurityalliance.org
NEAR WINDOW (Jul 20, 2026), flagged; zero duplicate hits. The Cloud Security Alliance AI Safety Initiative's independent post-mortem, the closest thing to a neutral third-party incident analysis given that most of the technical record is self-reported by the two companies involved. Quantifies the agentic phase at over ...
NEAR WINDOW (Jul 20, 2026), flagged; zero duplicate hits. The Cloud Security Alliance AI Safety Initiative's independent post-mortem, the closest thing to a neutral third-party incident analysis given that most of the technical record is self-reported by the two companies involved. Quantifies the agentic phase at over ...
NEAR WINDOW (Jul 20, 2026), flagged; zero duplicate hits. The Cloud Security Alliance AI Safety Initiative's independent post-mortem, the closest thing to a neutral third-party incident analysis given that most of the technical record is self-reported by the two companies involved. Quantifies the agentic phase at over ...
Context is managed as durable loop state rather than a single prompt payload. NEAR WINDOW (Jul 20, 2026), flagged; zero duplicate hits. The Cloud Security Alliance AI Safety Initiative's independent post-mortem, the closest thing to a neutral third-party incident analysis given that most of the technical record is self...
Use CSA Research Note: Hugging Face's Autonomous AI Agent Breach to bound risk before recurring or unattended execution.
Contextual source from labs.cloudsecurityalliance.org; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,138
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1138
2026-07-28
Govern
govern
Bound permissions, cost, failure, and escalation.
trigger;context;verification
operator;security
operations
enabling
practitioner-analysis
B
ok
https://labs.cloudsecurityalliance.org/research/csa-research-note-huggingface-autonomous-agent-breach-202607/
Hugging Face’s Autonomous AI Agent Breach – Lab Space
Key Takeaways Hugging Face disclosed on July 16, 2026 that an intrusion into its production infrastructure was driven end-to-end by an autonomous AI agent rather than a human operator at the keyboard, a scenario the company itself described as unlike anything it had previously handled [1][2].
2026-07-20
2026
Lab Space
html-meta
2026-07-29T08:06:54
ale-0481
Securing Unattended Loops
securing-unattended-loops
Blog
📝
Computer-Use and TOCTOU: What You Click Is Not What You Get
https://embracethered.com/blog/posts/2026/toctou-agent-what-you-click-is-not-what-you-get/
external
embracethered.com
OUTSIDE WINDOW (Jun 25, 2026), flagged for the maintainer to accept or reject; included only because the duplicate grep returns zero hits on embracethered.com entirely and the failure mode is structural to any perceive-then-act loop rather than model-specific. Johann Rehberger demonstrates a time-of-check/time-of-use r...
OUTSIDE WINDOW (Jun 25, 2026), flagged for the maintainer to accept or reject; included only because the duplicate grep returns zero hits on embracethered.com entirely and the failure mode is structural to any perceive-then-act loop rather than model-specific. Johann Rehberger demonstrates a time-of-check/time-of-use r...
OUTSIDE WINDOW (Jun 25, 2026), flagged for the maintainer to accept or reject; included only because the duplicate grep returns zero hits on embracethered.com entirely and the failure mode is structural to any perceive-then-act loop rather than model-specific. Johann Rehberger demonstrates a time-of-check/time-of-use r...
Untrusted intake is treated as a loop-level security boundary. OUTSIDE WINDOW (Jun 25, 2026), flagged for the maintainer to accept or reject; included only because the duplicate grep returns zero hits on embracethered.com entirely and the failure mode is structural to any perceive-then-act loop rather than model-specif...
Use Computer-Use and TOCTOU: What You Click Is Not What You Get to bound risk before recurring or unattended execution.
Contextual source from embracethered.com; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,139
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1139
2026-07-28
Govern
govern
Bound permissions, cost, failure, and escalation.
verification
operator;security
operations
enabling
practitioner-analysis
B
ok
https://embracethered.com/blog/posts/2026/toctou-agent-what-you-click-is-not-what-you-get/
Computer-Use and TOCTOU: What You Click Is Not What You Get! · Embrace The Red
Last year, Jun Kokatsu disclosed an interesting vulnerability with ChatGPT Operator by exploiting a race condition. I was wondering if I could reproduce this …
2026
Embrace The Red
html-meta
2026-07-29T08:06:54
ale-0482
Securing Unattended Loops
securing-unattended-loops
Benchmark
🧪
ContainmentBench: Trace-Based Evaluation of Post-Injection Containment in Tool-Using LLM Agents
https://arxiv.org/abs/2607.23999
external
arxiv.org
Argues prompt-injection evals that report only terminal attack/policy outcomes hide what actually happened after exposure. Separately measures endpoint policy compliance, logged propagation, recovery instrumentation, and authorized structured-action completion. Striking result from a pre-specified 17,640-rollout study:...
Argues prompt-injection evals that report only terminal attack/policy outcomes hide what actually happened after exposure. Separately measures endpoint policy compliance, logged propagation, recovery instrumentation, and authorized structured-action completion. Striking result from a pre-specified 17,640-rollout study:...
Argues prompt-injection evals that report only terminal attack/policy outcomes hide what actually happened after exposure. Separately measures endpoint policy compliance, logged propagation, recovery instrumentation, and authorized structured-action completion. Striking result from a pre-specified 17,640-rollout study:...
Evaluation data is used as the feedback signal for improving loop behavior. Argues prompt-injection evals that report only terminal attack/policy outcomes hide what actually happened after exposure. Separately measures endpoint policy compliance, logged propagation, recovery instrumentation, and authorized structured-a...
Use ContainmentBench: Trace-Based Evaluation of Post-Injection Containment in Tool-Using LLM Agents to bound risk before recurring or unattended execution.
Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope.
high
README.md
1,140
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1140
2026-07-28
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;verification;exit
researcher;evaluator;operator;security
operations
enabling
benchmark
A
ok
https://arxiv.org/abs/2607.23999
[2607.23999] ContainmentBench: Trace-Based Evaluation of Post-Injection Containment in Tool-Using LLM Agents
Tool-using LLM agents process untrusted content, maintain memory, delegate across agents, and invoke side-effecting tools. Existing prompt-injection evaluations typically summarize security with terminal attack or policy outcomes, but equal endpoints can conceal different post-exposure traces and different losses of au...
Wenhao Lan; Shan Li; Xinhua Lai; Meiqi Wu; Junbin Yang; Haihua Shen
2026-07-27
2026
arXiv
arXiv
28 pages, 7 figures
cs.CR
arxiv-api
2607.23999
2026-07-29T08:06:54
ale-0483
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Agentic Permissions Policy Algebra for Taint Confinement in LLM Agents
https://arxiv.org/abs/2607.24625
external
arxiv.org
APPA fixes the usability collapse of classic taint tracking, where reading one unvetted document permanently poisons the agent's context. Uses engine-managed context branching -- a label-seeded child trajectory absorbs the label descent locally and a trusted sanitizer returns a bounded result -- plus prospective acquis...
APPA fixes the usability collapse of classic taint tracking, where reading one unvetted document permanently poisons the agent's context. Uses engine-managed context branching -- a label-seeded child trajectory absorbs the label descent locally and a trusted sanitizer returns a bounded result -- plus prospective acquis...
APPA fixes the usability collapse of classic taint tracking, where reading one unvetted document permanently poisons the agent's context. Uses engine-managed context branching -- a label-seeded child trajectory absorbs the label descent locally and a trusted sanitizer returns a bounded result -- plus prospective acquis...
Context is managed as durable loop state rather than a single prompt payload. APPA fixes the usability collapse of classic taint tracking, where reading one unvetted document permanently poisons the agent's context. Uses engine-managed context branching -- a label-seeded child trajectory absorbs the label descent local...
Use Agentic Permissions Policy Algebra for Taint Confinement in LLM Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.24625; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,141
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1141
2026-07-28
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;context;exit
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.24625
[2607.24625] Agentic Permissions Policy Algebra for Taint Confinement in LLM Agents
Autonomous LLM agents processing mixed-confidentiality data face severe security risks from prompt injection attacks and reasoning errors. While dynamic Information Flow Control (IFC) provides structural security guarantees, traditional taint tracking permanently taints an agent's context upon reading unvetted data, se...
Arseny Kravchenko; Vadim Liventsev; Innokentii Konstantinov; Ildar Iskhakov; Matvey Kukuy
2026-07-27
2026
arXiv
arXiv
Preprint. Submitted to the 19th ACM Workshop on Artificial Intelligence and Security (AISec '26). 10 pages, 2 tables, 1 figure
cs.CR
arxiv-api
2607.24625
2026-07-29T08:06:54
ale-0484
Securing Unattended Loops
securing-unattended-loops
Paper
📄
What Can Be Enforced? A Theory of Certified Runtime Safety for Tool-Using Agents
https://arxiv.org/abs/2607.22868
external
arxiv.org
Theory paper on the limits of runtime guardrails that sit in front of irreversible tool calls. Separates three questions: what policies a deterministic gate can enforce given its register model (nontriviality undecidable with two decrementable counters, PSPACE for a separable monotone fragment); the exact false-block/m...
Theory paper on the limits of runtime guardrails that sit in front of irreversible tool calls. Separates three questions: what policies a deterministic gate can enforce given its register model (nontriviality undecidable with two decrementable counters, PSPACE for a separable monotone fragment); the exact false-block/m...
Theory paper on the limits of runtime guardrails that sit in front of irreversible tool calls. Separates three questions: what policies a deterministic gate can enforce given its register model (nontriviality undecidable with two decrementable counters, PSPACE for a separable monotone fragment); the exact false-block/m...
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Theory paper on the limits of runtime guardrails that sit in front of irreversible tool calls. Separates three questions: what policies a deterministic gate can enforce given its register model (nontriviality undecidable with two ...
Use What Can Be Enforced? A Theory of Certified Runtime Safety for Tool-Using Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.22868; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,142
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1142
2026-07-28
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.22868
[2607.22868] What Can Be Enforced? A Theory of Certified Runtime Safety for Tool-Using Agents
Runtime guardrails act before irreversible tool calls, but their guarantees depend on what policy state is representable, what a judge observes, and whether intervention changes future behavior. We separate three questions. First, relative to fixed oracle predicates, a deterministic gate enforces exactly the nonempty s...
Shawn Ray
2026-07-24
2026
arXiv
arXiv
26 pages, 8 figures. Extended version with complete proofs and additional experiments
cs.AI
arxiv-api
2607.22868
2026-07-29T08:06:54
ale-0485
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Are You Still the Agent I Authorized? Earned Authority under a Fixed Ceiling for Evolving Agents
https://arxiv.org/abs/2607.23586
external
arxiv.org
Formulates authorization continuity: long-lived agents evolve after deployment by retaining experience, acquiring skills and tools, revising workflows, and delegating, which can change both the effects reachable under an old grant and the authority a task now requires -- upward, downward, or incomparably. Existing tool...
Formulates authorization continuity: long-lived agents evolve after deployment by retaining experience, acquiring skills and tools, revising workflows, and delegating, which can change both the effects reachable under an old grant and the authority a task now requires -- upward, downward, or incomparably. Existing tool...
Formulates authorization continuity: long-lived agents evolve after deployment by retaining experience, acquiring skills and tools, revising workflows, and delegating, which can change both the effects reachable under an old grant and the authority a task now requires -- upward, downward, or incomparably. Existing tool...
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Formulates authorization continuity: long-lived agents evolve after deployment by retaining experience, acquiring skills and tools, revising workflows, and delegating, which can change both the effects reachable under an old grant...
Use Are You Still the Agent I Authorized? Earned Authority under a Fixed Ceiling for Evolving Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.23586; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,143
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1143
2026-07-28
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.23586
[2607.23586] Are You Still the Agent I Authorized? Earned Authority under a Fixed Ceiling for Evolving Agents
Long-lived AI agents increasingly evolve after deployment by retaining experience, acquiring skills and tools, revising workflows, delegating work, and moving across task phases. This improves adaptation but creates a distinct authorization problem. Tool-enabled agents can turn model errors and prompt injections into c...
Zhaoxi Zhang; Xiaomei Zhang
2026-07-26
2026
arXiv
arXiv
cs.AI
arxiv-api
2607.23586
2026-07-29T08:06:54
ale-0486
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Isolated but Exposed: Persistence-Based Memory Extraction Attack on LLM Agents
https://arxiv.org/abs/2607.23444
external
arxiv.org
Production long-term-memory systems bind each user's store to a unique identifier and assume isolation is sufficient. This shows the tool interface is the overlooked surface: agents routinely embed LTM-retrieved data in tool invocation parameters, so a malicious tool exfiltrates private memory without ever violating us...
Production long-term-memory systems bind each user's store to a unique identifier and assume isolation is sufficient. This shows the tool interface is the overlooked surface: agents routinely embed LTM-retrieved data in tool invocation parameters, so a malicious tool exfiltrates private memory without ever violating us...
Production long-term-memory systems bind each user's store to a unique identifier and assume isolation is sufficient. This shows the tool interface is the overlooked surface: agents routinely embed LTM-retrieved data in tool invocation parameters, so a malicious tool exfiltrates private memory without ever violating us...
Persistent memory is treated as an external runtime artifact. Production long-term-memory systems bind each user's store to a unique identifier and assume isolation is sufficient. This shows the tool interface is the overlooked surface: agents routinely embed LTM-retrieved data in tool invocation parameters, so a malic...
Use Isolated but Exposed: Persistence-Based Memory Extraction Attack on LLM Agents to bound risk before recurring or unattended execution.
Research source arXiv:2607.23444; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,144
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1144
2026-07-28
Govern
govern
Bound permissions, cost, failure, and escalation.
trigger;workspace;context;state;budget
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.23444
[2607.23444] Isolated but Exposed: Persistence-Based Memory Extraction Attack on LLM Agents
LLM-based agents extend large language models with long-term memory (LTM) that persists privacy-sensitive user data across sessions. Production systems mitigate extraction risks through memory isolation, binding each user's LTM to a unique identifier. This defense has blocked known attacks on shared storage, fostering ...
Xinyu Gao; Wenyu Chen; Xiangtao Meng; Li Wang; Chuanchao Zang; Jianing Wang; Zheng Li; Shanqing Guo
2026-07-26
2026
arXiv
arXiv
cs.CR
arxiv-api
2607.23444
2026-07-29T08:06:54
ale-0487
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Agent Security Needs Redefinition through a Holistic Framework
https://arxiv.org/abs/2607.22024
external
arxiv.org
Position paper arguing agent security is systematically misdefined as a question about action content -- does this instruction look malicious -- when it is fundamentally contextual. 'Delete user data' is either routine administration or a prompt injection, and content alone cannot separate them; authorization context c...
Position paper arguing agent security is systematically misdefined as a question about action content -- does this instruction look malicious -- when it is fundamentally contextual. 'Delete user data' is either routine administration or a prompt injection, and content alone cannot separate them; authorization context c...
Position paper arguing agent security is systematically misdefined as a question about action content -- does this instruction look malicious -- when it is fundamentally contextual. 'Delete user data' is either routine administration or a prompt injection, and content alone cannot separate them; authorization context c...
Context is managed as durable loop state rather than a single prompt payload. Position paper arguing agent security is systematically misdefined as a question about action content -- does this instruction look malicious -- when it is fundamentally contextual. 'Delete user data' is either routine administration or a pro...
Use Agent Security Needs Redefinition through a Holistic Framework to bound risk before recurring or unattended execution.
Research source arXiv:2607.22024; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,145
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1145
2026-07-28
Govern
govern
Bound permissions, cost, failure, and escalation.
context
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.22024
[2607.22024] Agent Security Needs Redefinition through a Holistic Framework
Agent security is widely treated as a question about action content. Defenses ask whether an instruction looks malicious. Benchmarks ask whether an agent performs a harmful sounding action. \textbf{We argue that agent security is fundamentally a contextual problem, and that the current content based framing systematica...
Vincent Siu; Jingxuan He; Kyle Montgomery; Zhun Wang; Chenguang Wang; Dawn Song
2026-07-24
2026
arXiv
arXiv
ICML 2026 Position Paper
cs.CR
arxiv-api
2607.22024
2026-07-29T08:06:54
ale-0488
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Separating Capability from Permission: A Governance Framework for Agentic AI Autonomy Levels
https://arxiv.org/abs/2607.23438
external
arxiv.org
Splits two things autonomy discussions routinely conflate: Autonomous Capability Levels (what the agent can technically do) and Allowed Autonomy Levels (what it is authorized to do given risk, oversight, and accountability). Defines levels spanning reactive execution, decision support, supervised action, goal-directed ...
Splits two things autonomy discussions routinely conflate: Autonomous Capability Levels (what the agent can technically do) and Allowed Autonomy Levels (what it is authorized to do given risk, oversight, and accountability). Defines levels spanning reactive execution, decision support, supervised action, goal-directed ...
Splits two things autonomy discussions routinely conflate: Autonomous Capability Levels (what the agent can technically do) and Allowed Autonomy Levels (what it is authorized to do given risk, oversight, and accountability). Defines levels spanning reactive execution, decision support, supervised action, goal-directed ...
Frames security as a recurring-loop boundary rather than a one-time prompt hygiene issue. Splits two things autonomy discussions routinely conflate: Autonomous Capability Levels (what the agent can technically do) and Allowed Autonomy Levels (what it is authorized to do given risk, oversight, and accountability). Defin...
Use Separating Capability from Permission: A Governance Framework for Agentic AI Autonomy Levels to bound risk before recurring or unattended execution.
Research source arXiv:2607.23438; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,146
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1146
2026-07-28
Govern
govern
Bound permissions, cost, failure, and escalation.
objective;workspace;delegation
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.23438
[2607.23438] Separating Capability from Permission: A Governance Framework for Agentic AI Autonomy Levels
As AI systems increasingly exhibit agentic behavior, discussions of autonomy often conflate what systems are technically capable of doing with what they should be permitted to do in practice. This paper introduces a governance framework that explicitly separates Allowed Autonomy Levels (AAL), which define the degree of...
Haining Zheng; Qian Dong; Rodolfo K. Depena; Jonathan D. Bhatia; Feng Xiao; Peng Xu
2026-07-26
2026
arXiv
arXiv
10 pages, 2 tables, 3 figures
cs.AI
arxiv-api
2607.23438
2026-07-29T08:06:54
ale-0489
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Dynamic Capability Scoping for Enterprise AI Agents: A Synthetic Dataset and Three-Source Permission Architecture
https://arxiv.org/abs/2607.22445
external
arxiv.org
Attacks standing over-privilege: enterprise agents get a static credential set at configuration time holding every tool the role might ever need, expanding the attack surface permanently. Argues capability scoping should be dynamic least-privilege and prevention-first, since a credential absent from context cannot be m...
Attacks standing over-privilege: enterprise agents get a static credential set at configuration time holding every tool the role might ever need, expanding the attack surface permanently. Argues capability scoping should be dynamic least-privilege and prevention-first, since a credential absent from context cannot be m...
Attacks standing over-privilege: enterprise agents get a static credential set at configuration time holding every tool the role might ever need, expanding the attack surface permanently. Argues capability scoping should be dynamic least-privilege and prevention-first, since a credential absent from context cannot be m...
Packages the evidence as queryable CSV and JSONL rather than only a rendered page. Attacks standing over-privilege: enterprise agents get a static credential set at configuration time holding every tool the role might ever need, expanding the attack surface permanently. Argues capability scoping should be dynamic least...
Use Dynamic Capability Scoping for Enterprise AI Agents: A Synthetic Dataset and Three-Source Permission Architecture to bound risk before recurring or unattended execution.
Research source arXiv:2607.22445; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,147
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1147
2026-07-28
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace;context
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.22445
[2607.22445] Dynamic Capability Scoping for Enterprise AI Agents: A Synthetic Dataset and Three-Source Permission Architecture
Enterprise AI agents are typically granted static credential sets at configuration time, holding every tool the role might need for every task they perform. This persistent over-privilege expands the attack surface. We argue that capability scoping must follow a dynamic least-privilege principle and be treated as a pre...
Halil Burak Noyan
2026-07-24
2026
arXiv
arXiv
Published at the Second Workshop on Agents in the Wild: Safety, Security, and Beyond (AIWILD) at ICML 2026
cs.AI
arxiv-api
2607.22445
2026-07-29T08:06:54
ale-0490
Securing Unattended Loops
securing-unattended-loops
Paper
📄
False Prophets: On the Security of World Models in Agentic Systems
https://arxiv.org/abs/2607.23147
external
arxiv.org
Agents increasingly use trained environment simulators to predict action outcomes before executing them -- a pattern that improves reliability but creates a new attack surface, since a manipulated prediction can steer the agent into harmful actions. Identifies world-model-specific vulnerabilities exploitable in termina...
Agents increasingly use trained environment simulators to predict action outcomes before executing them -- a pattern that improves reliability but creates a new attack surface, since a manipulated prediction can steer the agent into harmful actions. Identifies world-model-specific vulnerabilities exploitable in termina...
Agents increasingly use trained environment simulators to predict action outcomes before executing them -- a pattern that improves reliability but creates a new attack surface, since a manipulated prediction can steer the agent into harmful actions. Identifies world-model-specific vulnerabilities exploitable in termina...
Packages the evidence as queryable CSV and JSONL rather than only a rendered page. Agents increasingly use trained environment simulators to predict action outcomes before executing them -- a pattern that improves reliability but creates a new attack surface, since a manipulated prediction can steer the agent into harm...
Use False Prophets: On the Security of World Models in Agentic Systems to bound risk before recurring or unattended execution.
Research source arXiv:2607.23147; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,148
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1148
2026-07-28
Govern
govern
Bound permissions, cost, failure, and escalation.
verification
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.23147
[2607.23147] False Prophets: On the Security of World Models in Agentic Systems
Large language models now power autonomous agents capable of complex, multi-step tasks in different environments. Accurate and reliable execution of these tasks requires the agent to predict the results of its actions. Recent research proposes to enhance predictive capabilities via specially trained environment simulat...
Erik Imgrund; Anna Wimbauer; Klim Kireev; Konrad Rieck
2026-07-25
2026
arXiv
arXiv
cs.CR
arxiv-api
2607.23147
2026-07-29T08:06:54
ale-0491
Securing Unattended Loops
securing-unattended-loops
Paper
📄
Recursive Governance: A Graph-Theoretic Framework for Risk Propagation and Drift Detection in Agentic AI Systems
https://arxiv.org/abs/2607.23916
external
arxiv.org
Argues static model-inventory requirements from traditional model risk management become structurally obsolete once institutions run autonomous agentic systems, and proposes an Inventory-as-Code governance loop treating the inventory as a living architectural component rather than periodic documentation. Contributes a ...
Argues static model-inventory requirements from traditional model risk management become structurally obsolete once institutions run autonomous agentic systems, and proposes an Inventory-as-Code governance loop treating the inventory as a living architectural component rather than periodic documentation. Contributes a ...
Argues static model-inventory requirements from traditional model risk management become structurally obsolete once institutions run autonomous agentic systems, and proposes an Inventory-as-Code governance loop treating the inventory as a living architectural component rather than periodic documentation. Contributes a ...
Control flow is represented as an inspectable graph rather than an opaque prompt loop. Argues static model-inventory requirements from traditional model risk management become structurally obsolete once institutions run autonomous agentic systems, and proposes an Inventory-as-Code governance loop treating the inventory...
Use Recursive Governance: A Graph-Theoretic Framework for Risk Propagation and Drift Detection in Agentic AI Systems to bound risk before recurring or unattended execution.
Research source arXiv:2607.23916; inspect its method and evaluation before treating results as production evidence.
medium
README.md
1,149
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1149
2026-07-28
Govern
govern
Bound permissions, cost, failure, and escalation.
workspace
researcher;evaluator;operator;security
operations
enabling
research-preprint
A
ok
https://arxiv.org/abs/2607.23916
[2607.23916] Recursive Governance: A Graph-Theoretic Framework for Risk Propagation and Drift Detection in Agentic AI Systems
As financial institutions transition from traditional predictive models to autonomous agentic systems, the static model inventory requirements of traditional model risk management (MRM) face structural obsolescence. This paper proposes a dynamic Inventory-as-Code (IaC) governance loop that treats the model inventory as...
Sriram Nagaraj; Advaith Nila Narayanan
2026-07-27
2026
arXiv
arXiv
math.NA
arxiv-api
2607.23916
2026-07-29T08:06:54
ale-0492
State, Memory, And Context Persistence
state-memory-and-context-persistence
Docs
📚
Effective Context Engineering for AI Agents
https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
external
www.anthropic.com
Anthropic guide to context as managed runtime state rather than a prompt dump.
Anthropic guide to context as managed runtime state rather than a prompt dump.
Anthropic guide to context as managed runtime state rather than a prompt dump.
Context is managed as durable loop state rather than a single prompt payload. Anthropic guide to context as managed runtime state rather than a prompt dump.
Use Effective Context Engineering for AI Agents to carry context, state, and receipts across runs and failures.
Primary documentation from a platform, SDK, standard, or framework; strong implementation signal.
high
README.md
1,157
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1157
Persist
persist
Carry context, state, and receipts across runs.
context;state
builder
harness
enabling
technical-documentation
A
ok
https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
Effective context engineering for AI agents \ Anthropic
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
Anthropic
domain-fallback
2026-07-29T08:06:54
ale-0493
State, Memory, And Context Persistence
state-memory-and-context-persistence
Blog
📝
Agent Harnesses: the Infrastructure Layer Your LLM Agent Actually Needs
https://ninadpathak.com/blog/agent-harnesses/
external
ninadpathak.com
Covers execution loops, state, checkpointing, observers, and replayability.
Covers execution loops, state, checkpointing, observers, and replayability.
Covers execution loops, state, checkpointing, observers, and replayability.
Checkpointed state makes long-running agent work recoverable across failures. Covers execution loops, state, checkpointing, observers, and replayability.
Use Agent Harnesses: the Infrastructure Layer Your LLM Agent Actually Needs to carry context, state, and receipts across runs and failures.
Contextual source from ninadpathak.com; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,158
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1158
Persist
persist
Carry context, state, and receipts across runs.
state
builder
harness
enabling
practitioner-analysis
B
ok
https://ninadpathak.com/blog/agent-harnesses/
Agent Harnesses: the Infrastructure Layer Your Llm Agent Actually Needs | Ninad Pathak
Every production AI agent needs a harness. Here is what one contains, why frameworks often are not enough, and how to build the layer that actually determines reliability.
ninadpathak.com
domain-fallback
2026-07-29T08:06:54
ale-0494
State, Memory, And Context Persistence
state-memory-and-context-persistence
Blog
📝
The Agent Loop Is the New OS
https://www.harness.io/blog/agent-loop-new-os
external
www.harness.io
Frames the agent loop as an OS-like boundary with context as RAM and tools as I/O.
Frames the agent loop as an OS-like boundary with context as RAM and tools as I/O.
Frames the agent loop as an OS-like boundary with context as RAM and tools as I/O.
Context is managed as durable loop state rather than a single prompt payload. Frames the agent loop as an OS-like boundary with context as RAM and tools as I/O.
Use The Agent Loop Is the New OS to carry context, state, and receipts across runs and failures.
Contextual source from www.harness.io; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,159
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1159
Persist
persist
Carry context, state, and receipts across runs.
workspace;context
builder
harness
enabling
practitioner-analysis
B
ok
https://www.harness.io/blog/agent-loop-new-os
The Agent Loop Is the New OS | Harness Blog | Harness Share in X Share in Facebook Share in LinkedIn Search in ChatGpt Github icon LinkedIn icon Facebook icon Instagram icon Twitter icon
The Harness MCP server treats the AI agent loop as an operating system, mapping the LLM to the CPU and the Context Window to RAM. Learn how this design uses 10 generic, composable tools to abstract complexity and keep the context window clean for higher-quality, cost-efficient AI agent reasoning. | Blog
2026
2026
Harness.io
html-meta
2026-07-29T08:06:54
ale-0495
State, Memory, And Context Persistence
state-memory-and-context-persistence
Blog
📝
Harness engineering for coding agent users
https://martinfowler.com/articles/harness-engineering.html
external
martinfowler.com
Martin Fowler article on feedforward, feedback, and outer harnesses for coding agents.
Martin Fowler article on feedforward, feedback, and outer harnesses for coding agents.
Martin Fowler article on feedforward, feedback, and outer harnesses for coding agents.
Makes persistence and context management visible as runtime design choices. Martin Fowler article on feedforward, feedback, and outer harnesses for coding agents.
Use Harness engineering for coding agent users to carry context, state, and receipts across runs and failures.
Contextual source from martinfowler.com; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,160
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1160
Persist
persist
Carry context, state, and receipts across runs.
context;state
builder
harness
enabling
practitioner-analysis
B
ok
https://martinfowler.com/articles/harness-engineering.html
Harness engineering for coding agent users
A mental model for building trust in coding agents through feedforward guides, feedback sensors, and iterative harness engineering.
martinfowler.com
html-meta
2026-07-29T08:06:54
ale-0496
State, Memory, And Context Persistence
state-memory-and-context-persistence
Blog
📝
Context Engineering
https://simonwillison.net/2025/Jun/27/context-engineering/
external
simonwillison.net
Simon Willison's framing of context engineering, useful for distinguishing context state from loop orchestration.
Simon Willison's framing of context engineering, useful for distinguishing context state from loop orchestration.
Simon Willison's framing of context engineering, useful for distinguishing context state from loop orchestration.
Context is managed as durable loop state rather than a single prompt payload. Simon Willison's framing of context engineering, useful for distinguishing context state from loop orchestration.
Use Context Engineering to carry context, state, and receipts across runs and failures.
Contextual source from simonwillison.net; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,161
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1161
Persist
persist
Carry context, state, and receipts across runs.
context;delegation;state
builder
harness
enabling
practitioner-analysis
B
ok
https://simonwillison.net/2025/Jun/27/context-engineering/
Context engineering
The term context engineering has recently started to gain traction as a better alternative to prompt engineering. I like it. I think this one may have sticking power. Here's an …
Simon Willison
2025
Simon Willison’s Weblog
html-meta
2026-07-29T08:06:54
ale-0497
State, Memory, And Context Persistence
state-memory-and-context-persistence
Blog
📝
Agentic Coding in 2026
https://sourcegraph.com/blog/agentic-coding
external
sourcegraph.com
Sourcegraph on supplying deterministic, large-codebase context and code intelligence so recurring agent runs reuse durable repository state instead of rediscovering it each time.
Sourcegraph on supplying deterministic, large-codebase context and code intelligence so recurring agent runs reuse durable repository state instead of rediscovering it each time.
Sourcegraph on supplying deterministic, large-codebase context and code intelligence so recurring agent runs reuse durable repository state instead of rediscovering it each time.
Durable execution and replay are treated as first-class loop infrastructure. Sourcegraph on supplying deterministic, large-codebase context and code intelligence so recurring agent runs reuse durable repository state instead of rediscovering it each time.
Use Agentic Coding in 2026 to carry context, state, and receipts across runs and failures.
Contextual source from sourcegraph.com; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,162
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1162
Persist
persist
Carry context, state, and receipts across runs.
context;state
builder
harness
enabling
practitioner-analysis
B
ok
https://sourcegraph.com/blog/agentic-coding
Agentic Coding in 2026: A Practical Guide for Big Code | Sourcegraph
Learn what agentic coding is, how AI coding agents work in real engineering orgs, and how to give them the codebase context they need to ship safely.
Sourcegraph
html-meta
2026-07-29T08:06:54
ale-0498
State, Memory, And Context Persistence
state-memory-and-context-persistence
Blog
📝
Agentic AI State Management with ScyllaDB and LangGraph
https://www.scylladb.com/2026/04/08/agentic-ai-state-management-with-scylladb-and-langgraph/
external
www.scylladb.com
Durable agent state with checkpointers, write-ahead logs, and time-travel branching.
Durable agent state with checkpointers, write-ahead logs, and time-travel branching.
Durable agent state with checkpointers, write-ahead logs, and time-travel branching.
Durable execution and replay are treated as first-class loop infrastructure. Durable agent state with checkpointers, write-ahead logs, and time-travel branching.
Use Agentic AI State Management with ScyllaDB and LangGraph to carry context, state, and receipts across runs and failures.
Contextual source from www.scylladb.com; useful for practice signals or boundary conditions, not independent validation.
contextual
README.md
1,163
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1163
Persist
persist
Carry context, state, and receipts across runs.
state
builder
harness
enabling
practitioner-analysis
B
ok
https://www.scylladb.com/2026/04/08/agentic-ai-state-management-with-scylladb-and-langgraph/
Agentic AI State Management with ScyllaDB and LangGraph | ScyllaDB
How to combine LangGraph and ScyllaDB for durable state management, crash recovery, and a highly available backend for your agentic AI applications.
Cynthia Dunlop
2026-04-08
2026
ScyllaDB
html-meta
2026-07-29T08:06:54
ale-0499
State, Memory, And Context Persistence
state-memory-and-context-persistence
Tool
🧰
Mem0
https://github.com/mem0ai/mem0
external
github.com
Open-source memory layer for retaining user, session, and agent state across repeated agent sessions.
Open-source memory layer for retaining user, session, and agent state across repeated agent sessions.
Open-source memory layer for retaining user, session, and agent state across repeated agent sessions.
Persistent memory is treated as an external runtime artifact. Open-source memory layer for retaining user, session, and agent state across repeated agent sessions.
Use Mem0 to carry context, state, and receipts across runs and failures.
Inspectable GitHub source (61,987 stars; 7,229 forks; Apache-2.0 license; updated 2026-07-29); popularity is context, not proof of reliability.
medium
README.md
1,164
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1164
Persist
persist
Carry context, state, and receipts across runs.
context;state
builder
harness
enabling
source-implementation
A
ok
https://github.com/mem0ai/mem0
GitHub - mem0ai/mem0: Universal memory layer for AI Agents · GitHub
Universal memory layer for AI Agents. Contribute to mem0ai/mem0 development by creating an account on GitHub.
2023-06-20
2023
mem0ai/mem0
GitHub
github-api
mem0ai/mem0
61987
7229
Apache-2.0
2023-06-20T08:58:36Z
2026-07-29T07:57:22Z
2026-07-29T08:06:54
ale-0500
State, Memory, And Context Persistence
state-memory-and-context-persistence
Tool
🧰
Letta
https://github.com/letta-ai/letta
external
github.com
Stateful agent framework from the MemGPT line with persistent, self-editing memory across runs.
Stateful agent framework from the MemGPT line with persistent, self-editing memory across runs.
Stateful agent framework from the MemGPT line with persistent, self-editing memory across runs.
Persistent memory is treated as an external runtime artifact. Stateful agent framework from the MemGPT line with persistent, self-editing memory across runs.
Use Letta to carry context, state, and receipts across runs and failures.
Inspectable GitHub source (24,008 stars; 2,560 forks; Apache-2.0 license; updated 2026-07-29); popularity is context, not proof of reliability.
medium
README.md
1,165
https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1165
Persist
persist
Carry context, state, and receipts across runs.
context;state
builder
harness
enabling
source-implementation
A
ok
https://github.com/letta-ai/letta
GitHub - letta-ai/letta: Platform for stateful agents: AI with advanced memory that can learn and self-improve over time. · GitHub
Platform for stateful agents: AI with advanced memory that can learn and self-improve over time. - letta-ai/letta
2023-10-11
2023
letta-ai/letta
GitHub
github-api
letta-ai/letta
24008
2560
Apache-2.0
2023-10-11T07:38:37Z
2026-07-29T07:35:25Z
2026-07-29T08:06:54