Daily Papers

Large language model (LLM) coding agents increasingly operate with shell access, allowing them to write scripts, move files, alter permissions, and modify system configurations. However, an agent may achieve a user's goal while causing unintended harm, such as deleting unrelated files, weakening access controls, or corrupting data, in response to benign inputs without explicit jailbreaking. This failure mode is particularly important because apparent task success may reassure users that an agent behaved correctly, causing harmful side effects to go unnoticed. We call this harmful success: the agent successfully completes the intended benign task while simultaneously causing unintended harm. Despite its importance, harmful success is difficult to measure because conventional task-success metrics overlook harmful behaviors, while existing safety evaluations are often designed solely to assess harmful outcomes. We introduce VeriHarm, an automatic framework for constructing programmatically verifiable benchmarks that directly measure harmful success. VeriHarm minimally perturbs coding tasks to elicit unintended harmful behavior while maintaining the benignity of the task and preserving the original task objective. It then synthesizes executable harm tests that, together with the original task-correctness tests, evaluate both task accuracy and harmful outcomes. Frontier agents, including GPT-5.6-Sol and Claude-Opus-5, achieve an average task accuracy of 95% on our curated VeriHarm-Bench, yet still struggle to complete tasks safely: their safe success rate, the fraction of tasks completed successfully without triggering the designated harm, averages only 37%. These results suggest that successful task completion can frequently coexist with unintended harm, and that task-success metrics alone may substantially understate the potential safety risks of coding agents.

Arabic harmful-language detection has received considerable attention, yet Arabic text detoxification remains underexplored. We introduce AraDetox, a multi-dialect Arabic detoxification dataset comprising 10,500 harmful social-media posts and 84,000 detoxified rewrites generated using GPT-5 and Gemini 2.5 Flash across Modern Standard Arabic, Gulf, Levantine, and Egyptian Arabic. The generated outputs were assessed through human evaluation and automatic analyses of lexical change, semantic preservation, sentiment, and dialectal style. Results show that detoxification is primarily a meaning-preserving rewriting task: substantial lexical and structural reformulation is accompanied by consistently high semantic similarity. Human evaluation confirms successful harmful-language removal while largely preserving the original meaning. Dialectal analyses further indicate that the generated variants exhibit measurable stylistic alignment with reference Arabic dialect corpora. Comparison with existing resources highlights two complementary approaches to detoxification: minimal-edit lexical substitution and meaning-preserving reformulation. Our findings demonstrate that large-scale Arabic detoxification resources can be constructed through LLM-assisted generation and human verification. The dataset is publicly available at https://github.com/ArabicNLP-UK/AraDetox to support future research on Arabic detoxification, safe text generation, and multi-dialect Arabic NLP.

To address the risks of encountering inappropriate or harmful content, researchers managed to incorporate several harmful contents datasets with machine learning methods to detect harmful concepts. However, existing harmful datasets are curated by the presence of a narrow range of harmful objects, and only cover real harmful content sources. This hinders the generalizability of methods based on such datasets, potentially leading to misjudgments. Therefore, we propose a comprehensive harmful dataset, Visual Harmful Dataset 11K (VHD11K), consisting of 10,000 images and 1,000 videos, crawled from the Internet and generated by 4 generative models, across a total of 10 harmful categories covering a full spectrum of harmful concepts with nontrivial definition. We also propose a novel annotation framework by formulating the annotation process as a multi-agent Visual Question Answering (VQA) task, having 3 different VLMs "debate" about whether the given image/video is harmful, and incorporating the in-context learning strategy in the debating process. Therefore, we can ensure that the VLMs consider the context of the given image/video and both sides of the arguments thoroughly before making decisions, further reducing the likelihood of misjudgments in edge cases. Evaluation and experimental results demonstrate that (1) the great alignment between the annotation from our novel annotation framework and those from human, ensuring the reliability of VHD11K; (2) our full-spectrum harmful dataset successfully identifies the inability of existing harmful content detection methods to detect extensive harmful contents and improves the performance of existing harmfulness recognition methods; (3) VHD11K outperforms the baseline dataset, SMID, as evidenced by the superior improvement in harmfulness recognition methods. The complete dataset and code can be found at https://github.com/nctu-eva-lab/VHD11K.

Text-to-image (T2I) diffusion models offer powerful visual generation, but their controllability creates a critical safety challenge: adversarial prompts can steer the denoising trajectory toward policy-violating content such as explicit nudity or graphic violence. Existing safeguards mostly act before generation through prompt filtering or after generation through image classification, leaving the diffusion process itself unguarded and often yielding only refusal rather than safe visual repair. We introduce GuardPaint, a speculative decoding framework for safe T2I generation that intervenes inside the diffusion trajectory without modifying the base model. A lightweight auditor monitors intermediate images, localizes unsafe regions, and triggers surgical inpainting repair only where needed. Candidate repairs are generated by a policy-aligned inpainter and selected through a guarded tournament that accepts edits only when they improve policy compliance while preserving prompt fidelity and perceptual quality. Across five jailbreak families SneakPrompt, MMA, PGJ, DACA, and RABell and UNet/flow-matching models including SD~1.5, SDXL, SD~3.5, and FLUX.1-dev. GuardPaint reduces attack success and harmful generations with minimal degradation to image quality, prompt fidelity, and benign behavior. Content warning: This paper contains examples involving nudity and violence that some readers may find disturbing, distressing, or offensive.

With the emergence of strong visual-language capabilities, multimodal large language models (MLLMs) have demonstrated tremendous potential for real-world applications. However, the security vulnerabilities exhibited by the visual modality pose significant challenges to deploying such models in open-world environments. Recent studies have successfully induced harmful responses from target MLLMs by encoding harmful textual semantics directly into visual inputs. However, in these approaches, the visual modality primarily serves as a trigger for unsafe behavior, often exhibiting semantic ambiguity and lacking grounding in realistic scenarios. In this work, we define a novel setting: visual-centric jailbreak, where visual information serves as a necessary component in constructing a complete and realistic jailbreak context. Building on this setting, we propose the VisCo (Visual Contextual) Attack. VisCo fabricates contextual dialogue using four distinct visual-focused strategies, dynamically generating auxiliary images when necessary to construct a visual-centric jailbreak scenario. To maximize attack effectiveness, it incorporates automatic toxicity obfuscation and semantic refinement to produce a final attack prompt that reliably triggers harmful responses from the target black-box MLLMs. Specifically, VisCo achieves a toxicity score of 4.78 and an Attack Success Rate (ASR) of 85% on MM-SafetyBench against GPT-4o, significantly outperforming the baseline, which performs a toxicity score of 2.48 and an ASR of 22.2%. The code is available at https://github.com/Dtc7w3PQ/Visco-Attack.

Numerous studies have investigated methods for jailbreaking Large Language Models (LLMs) to generate harmful content. Typically, these methods are evaluated using datasets of malicious prompts designed to bypass security policies established by LLM providers. However, the generally broad scope and open-ended nature of existing datasets can complicate the assessment of jailbreaking effectiveness, particularly in specific domains, notably cybersecurity. To address this issue, we present and publicly release CySecBench, a comprehensive dataset containing 12662 prompts specifically designed to evaluate jailbreaking techniques in the cybersecurity domain. The dataset is organized into 10 distinct attack-type categories, featuring close-ended prompts to enable a more consistent and accurate assessment of jailbreaking attempts. Furthermore, we detail our methodology for dataset generation and filtration, which can be adapted to create similar datasets in other domains. To demonstrate the utility of CySecBench, we propose and evaluate a jailbreaking approach based on prompt obfuscation. Our experimental results show that this method successfully elicits harmful content from commercial black-box LLMs, achieving Success Rates (SRs) of 65% with ChatGPT and 88% with Gemini; in contrast, Claude demonstrated greater resilience with a jailbreaking SR of 17%. Compared to existing benchmark approaches, our method shows superior performance, highlighting the value of domain-specific evaluation datasets for assessing LLM security measures. Moreover, when evaluated using prompts from a widely used dataset (i.e., AdvBench), it achieved an SR of 78.5%, higher than the state-of-the-art methods.

Recently, language models like Llama 3.1 Instruct have become increasingly capable of agentic behavior, enabling them to perform tasks requiring short-term planning and tool use. In this study, we apply refusal-vector ablation to Llama 3.1 70B and implement a simple agent scaffolding to create an unrestricted agent. Our findings imply that these refusal-vector ablated models can successfully complete harmful tasks, such as bribing officials or crafting phishing attacks, revealing significant vulnerabilities in current safety mechanisms. To further explore this, we introduce a small Safe Agent Benchmark, designed to test both harmful and benign tasks in agentic scenarios. Our results imply that safety fine-tuning in chat models does not generalize well to agentic behavior, as we find that Llama 3.1 Instruct models are willing to perform most harmful tasks without modifications. At the same time, these models will refuse to give advice on how to perform the same tasks when asked for a chat completion. This highlights the growing risk of misuse as models become more capable, underscoring the need for improved safety frameworks for language model agents.

Large language models are increasingly used for mental health support, yet their conversational coherence alone does not ensure clinical appropriateness. Existing general-purpose safeguards often fail to distinguish between therapeutic disclosures and genuine clinical crises, leading to safety failures. To address this gap, we introduce a clinically grounded risk taxonomy, developed in collaboration with PhD-level psychologists, that identifies actionable harm (e.g., self-harm and harm to others) while preserving space for safe, non-crisis therapeutic content. We release MindGuard-testset, a dataset of real-world multi-turn conversations annotated at the turn level by clinical experts. Using synthetic dialogues generated via a controlled two-agent setup, we train MindGuard, a family of lightweight safety classifiers (with 4B and 8B parameters). Our classifiers reduce false positives at high-recall operating points and, when paired with clinician language models, help achieve lower attack success and harmful engagement rates in adversarial multi-turn interactions compared to general-purpose safeguards. We release all models and human evaluation data.

As large language models (LLMs) are increasingly deployed in real-world applications, the need to selectively remove unwanted knowledge while preserving model utility has become paramount. Recent work has explored sparse autoencoders (SAEs) to perform precise interventions on monosemantic features. However, most SAE-based methods operate at inference time, which does not create persistent changes in the model's parameters. Such interventions can be bypassed or reversed by malicious actors with parameter access. We introduce CRISP, a parameter-efficient method for persistent concept unlearning using SAEs. CRISP automatically identifies salient SAE features across multiple layers and suppresses their activations. We experiment with two LLMs and show that our method outperforms prior approaches on safety-critical unlearning tasks from the WMDP benchmark, successfully removing harmful knowledge while preserving general and in-domain capabilities. Feature-level analysis reveals that CRISP achieves semantically coherent separation between target and benign concepts, allowing precise suppression of the target features.

LLM based agents are increasingly deployed in high stakes settings where they process external data sources such as emails, documents, and code repositories. This creates exposure to indirect prompt injection attacks, where adversarial instructions embedded in external content manipulate agent behavior without user awareness. A critical but underexplored dimension of this threat is concealment: since users tend to observe only an agent's final response, an attack can conceal its existence by presenting no clue of compromise in the final user facing response while successfully executing harmful actions. This leaves users unaware of the manipulation and likely to accept harmful outcomes as legitimate. We present findings from a large scale public red teaming competition evaluating this dual objective across three agent settings: tool calling, coding, and computer use. The competition attracted 464 participants who submitted 272000 attack attempts against 13 frontier models, yielding 8648 successful attacks across 41 scenarios. All models proved vulnerable, with attack success rates ranging from 0.5% (Claude Opus 4.5) to 8.5% (Gemini 2.5 Pro). We identify universal attack strategies that transfer across 21 of 41 behaviors and multiple model families, suggesting fundamental weaknesses in instruction following architectures. Capability and robustness showed weak correlation, with Gemini 2.5 Pro exhibiting both high capability and high vulnerability. To address benchmark saturation and obsoleteness, we will endeavor to deliver quarterly updates through continued red teaming competitions. We open source the competition environment for use in evaluations, along with 95 successful attacks against Qwen that did not transfer to any closed source model. We share model-specific attack data with respective frontier labs and the full dataset with the UK AISI and US CAISI to support robustness research.

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 scores an agent's tool-call trajectory on a seven-level ordinal scale (L0 to L6) according to whether the executed action was reversible, whether it crossed scope to reach another party, and whether it expanded privilege. We compute the scale two ways: a deterministic oracle that reads the trajectory and the attacker's stated goal, and a panel of three frontier language-model judges that read a tag-free account of the same trajectory. Across four victim models and two defenses on the AgentDojo workspace suite, severity grading exposes three cases the binary metric hides, including a defense that reports a zero attack-success rate while still permitting an externally visible cross-scope leak through an unfiltered tool. The judge panel reproduces the oracle with high ordinal agreement (Krippendorff's alpha = 0.91) but shares systematic blind spots that we characterize, most notably a failure to recognize escalation chains. Unlike prior work that provides harm taxonomies, harmful-task completion tests, execution-level safety benchmarks, or severity-aware simulation, our contribution is a reusable, trace-grounded severity instrument applied to the actual actions recorded in existing red-team logs. All code, prompts, and per-episode logs are released.

Agent skills are the de facto mechanism for extending LLM agents with reusable guidance. A skill can shape the agent's task execution, including planning, tool use, problem-solving, and validation. Prior work reported mixed results of agent skills: some skills improve task success rates, while others have no effect, increase token use and execution time, and even reduce success rates. This paper presents a comprehensive analysis of skill-induced agent failures by attributing task failures and cost regressions to specific loaded skills. We introduce a differential analysis framework that attributes a failure or regression to a skill by comparing a target skill-guided run against a no-skill or semantically matched skill reference run that solves the same task, or solves it more cheaply. We instantiate this framework on SkillsBench and SWE-Skills-Bench, yielding 307 skill-induced failures, including 125 functional failures and 182 efficiency regressions. We also build SkillTriage, a taxonomy-guided attribution tool that normalizes paired cases, extracts differential evidence, and produces triage reports. Our major findings include: (1) Skill induced functional failures are rarely caused by obviously irrelevant skills; instead, seemingly relevant skills often make the agent incorrectly implement or omit task-required implementation elements. (2) Skill-induced efficiency regressions are not explained by prompt length alone. (3) The largest sources within Excessive Procedure are excessive verification and heavy implementation pipelines, contributing 67 and 30 cases, respectively. This shows that skills often turn validation checklists and construction recipes into mandatory work. Based on our findings, we propose research topics and tooling improvements for safer and more cost-aware skill reuse.

While Large Language Models (LLMs) have demonstrated impressive performance in various domains and tasks, concerns about their safety are becoming increasingly severe. In particular, since models may store unsafe knowledge internally, machine unlearning has emerged as a representative paradigm to ensure model safety. Existing approaches employ various training techniques, such as gradient ascent and negative preference optimization, in attempts to eliminate the influence of undesired data on target models. However, these methods merely suppress the activation of undesired data through parametric training without completely eradicating its informational traces within the model. This fundamental limitation makes it difficult to achieve effective continuous unlearning, rendering these methods vulnerable to relearning attacks. To overcome these challenges, we propose a Metamorphosis Representation Projection (MRP) approach that pioneers the application of irreversible projection properties to machine unlearning. By implementing projective transformations in the hidden state space of specific network layers, our method effectively eliminates harmful information while preserving useful knowledge. Experimental results demonstrate that our approach enables effective continuous unlearning and successfully defends against relearning attacks, achieving state-of-the-art performance in unlearning effectiveness while preserving natural performance. Our code is available in https://github.com/ChengcanWu/MRP.

Jailbreaking attacks can enable Large Language Models (LLMs) to bypass the safeguard and generate harmful content. Existing jailbreaking defense methods have failed to address the fundamental issue that harmful knowledge resides within the model, leading to potential jailbreak risks for LLMs. In this paper, we propose a novel defense method called Eraser, which mainly includes three goals: unlearning harmful knowledge, retaining general knowledge, and maintaining safety alignment. The intuition is that if an LLM forgets the specific knowledge required to answer a harmful question, it will no longer have the ability to answer harmful questions. The training of Erase does not actually require the model's own harmful knowledge, and it can benefit from unlearning general answers related to harmful queries, which means it does not need assistance from the red team. The experimental results show that Eraser can significantly reduce the jailbreaking success rate for various attacks without compromising the general capabilities of the model. Our codes are available at https://github.com/ZeroNLP/Eraser.

Large language models (LLMs) demonstrate impressive capabilities in various language tasks but are susceptible to jailbreak attacks that circumvent their safety alignments. This paper introduces Latent Fusion Jailbreak (LFJ), a representation-based attack that interpolates hidden states from harmful and benign query pairs to elicit prohibited responses. LFJ begins by selecting query pairs with high thematic and syntactic similarity, then performs gradient-guided interpolation at influential layers and tokens, followed by optimization to balance attack success, output fluency, and computational efficiency. Evaluations on models such as Vicuna and LLaMA-2 across benchmarks like AdvBench and MaliciousInstruct yield an average attack success rate (ASR) of 94.01%, outperforming existing methods. To mitigate LFJ, we propose an adversarial training defense that fine-tunes models on interpolated examples, reducing ASR by over 80% without degrading performance on benign inputs. Ablation studies validate the importance of query pair selection, hidden state interpolation components, and optimization strategies in LFJ's effectiveness.

Safety trained large language models (LLMs) can often be induced to answer harmful requests through jailbreak prompts. Because we lack a robust understanding of why LLMs are susceptible to jailbreaks, future frontier models operating more autonomously in higher-stakes settings may similarly be vulnerable to such attacks. Prior work has studied jailbreak success by examining the model's intermediate representations, identifying directions in this space that causally encode concepts like harmfulness and refusal. Then, they globally explain all jailbreak attacks as attempting to reduce or strengthen these concepts (e.g., reduce harmfulness). However, different jailbreak strategies may succeed by strengthening or suppressing different intermediate concepts, and the same jailbreak strategy may not work for different harmful request categories (e.g., violence vs. cyberattack); thus, we seek to give a local explanation -- i.e., why did this specific jailbreak succeed? To address this gap, we introduce LOCA, a method that gives Local, CAusal explanations of jailbreak success by identifying a minimal set of interpretable, intermediate representation changes that causally induce model refusal on an otherwise successful jailbreak request. We evaluate LOCA on harmful original-jailbreak pairs from a large jailbreak benchmark across Gemma and Llama chat models, comparing against prior methods adapted to this setting. LOCA can successfully induce refusal by making, on average, six interpretable changes; prior work routinely fails to achieve refusal even after 20 changes. LOCA is a step toward mechanistic, local explanations of jailbreak success in LLMs. Code to be released.

Skills are becoming the capability layer through which LLM agents turn plans into actions, but their use introduces security risks such as data leakage, unauthorized operations, and tool misuse. Existing vetting usually evaluates each skill in isolation, while real agent tasks often invoke multiple skills in a shared execution context. This creates Skill Composition Risk (SCR): a skill that appears benign alone can become harmful when its outputs, trust signals, authorization cues, or side effects influence later invocations along an activated path. We introduce SCR-Bench to evaluate this risk in controlled, sandboxed skill environments. Rather than relying only on textual intent or surface behavior, SCR-Bench records downstream state changes and path-level outcomes across composed skill executions. It contains three sub-benchmarks: SCR-CapFlow for capability-flow composition, SCR-TrustLift for trust-transfer composition, and SCR-AuthBlur for authorization-confusion composition. Across SCR-Bench, composed paths expose risks that are largely absent under isolated evaluation. In SCR-CapFlow, attack success rate reaches 33.6 percent under composition, compared with near-zero isolated baselines. In SCR-TrustLift, attack success rate exceeds 96.5 percent on four of five backends. In SCR-AuthBlur, the risky-approval rate increases by 71.8 percent relative to the L0 isolated baseline under the L1 context setting. These results show that agent skill security should be assessed at the level of activated paths rather than isolated artifacts. SCR and SCR-Bench provide a foundation for path-aware risk evaluation and defense in LLM agent skill ecosystems. Benchmark: https://github.com/saint-viperx/SCR_Bench.

Despite extensive safety alignment efforts, large language models (LLMs) remain vulnerable to jailbreak attacks that elicit harmful behavior. While existing studies predominantly focus on attack methods that require technical expertise, two critical questions remain underexplored: (1) Are jailbroken responses truly useful in enabling average users to carry out harmful actions? (2) Do safety vulnerabilities exist in more common, simple human-LLM interactions? In this paper, we demonstrate that LLM responses most effectively facilitate harmful actions when they are both actionable and informative--two attributes easily elicited in multi-step, multilingual interactions. Using this insight, we propose HarmScore, a jailbreak metric that measures how effectively an LLM response enables harmful actions, and Speak Easy, a simple multi-step, multilingual attack framework. Notably, by incorporating Speak Easy into direct request and jailbreak baselines, we see an average absolute increase of 0.319 in Attack Success Rate and 0.426 in HarmScore in both open-source and proprietary LLMs across four safety benchmarks. Our work reveals a critical yet often overlooked vulnerability: Malicious users can easily exploit common interaction patterns for harmful intentions.

Evaluating jailbreak attacks is challenging when prompts are not overtly harmful or fail to induce harmful outputs. Unfortunately, many existing red-teaming datasets contain such unsuitable prompts. To evaluate attacks accurately, these datasets need to be assessed and cleaned for maliciousness. However, existing malicious content detection methods rely on either manual annotation, which is labor-intensive, or large language models (LLMs), which have inconsistent accuracy in harmful types. To balance accuracy and efficiency, we propose a hybrid evaluation framework named MDH (Malicious content Detection based on LLMs with Human assistance) that combines LLM-based annotation with minimal human oversight, and apply it to dataset cleaning and detection of jailbroken responses. Furthermore, we find that well-crafted developer messages can significantly boost jailbreak success, leading us to propose two new strategies: D-Attack, which leverages context simulation, and DH-CoT, which incorporates hijacked chains of thought. The Codes, datasets, judgements, and detection results will be released in github repository: https://github.com/AlienZhang1996/DH-CoT.

Jailbreak attacks pose a serious threat to the safety of Large Language Models (LLMs) by crafting adversarial prompts that bypass alignment mechanisms, causing the models to produce harmful, restricted, or biased content. In this paper, we propose SafeLLM, a novel unlearning-based defense framework that unlearn the harmful knowledge from LLMs while preserving linguistic fluency and general capabilities. SafeLLM employs a three-stage pipeline: (1) dynamic unsafe output detection using a hybrid approach that integrates external classifiers with model-internal evaluations; (2) token-level harmful content tracing through feedforward network (FFN) activations to localize harmful knowledge; and (3) constrained optimization to suppress unsafe behavior without degrading overall model quality. SafeLLM achieves targeted and irreversible forgetting by identifying and neutralizing FFN substructures responsible for harmful generation pathways. Extensive experiments on prominent LLMs (Vicuna, LLaMA, and GPT-J) across multiple jailbreak benchmarks show that SafeLLM substantially reduces attack success rates while maintaining high general-purpose performance. Compared to standard defense methods such as supervised fine-tuning and direct preference optimization, SafeLLM offers stronger safety guarantees, more precise control over harmful behavior, and greater robustness to unseen attacks. Moreover, SafeLLM maintains the general performance after the harmful knowledge unlearned. These results highlight unlearning as a promising direction for scalable and effective LLM safety.

Large language models (LLMs) are increasingly deployed in contexts where their failures can have direct sociopolitical consequences. Yet, existing safety benchmarks rarely test vulnerabilities in domains such as political manipulation, propaganda and disinformation generation, or surveillance and information control. We introduce SocialHarmBench, a dataset of 585 prompts spanning 7 sociopolitical categories and 34 countries, designed to surface where LLMs most acutely fail in politically charged contexts. Our evaluations reveal several shortcomings: open-weight models exhibit high vulnerability to harmful compliance, with Mistral-7B reaching attack success rates as high as 97% to 98% in domains such as historical revisionism, propaganda, and political manipulation. Moreover, temporal and geographic analyses show that LLMs are most fragile when confronted with 21st-century or pre-20th-century contexts, and when responding to prompts tied to regions such as Latin America, the USA, and the UK. These findings demonstrate that current safeguards fail to generalize to high-stakes sociopolitical settings, exposing systematic biases and raising concerns about the reliability of LLMs in preserving human rights and democratic values. We share the SocialHarmBench benchmark at https://huggingface.co/datasets/psyonp/SocialHarmBench.

Computer-use agents extend language models from text generation to persistent action over tools, files, and execution environments. Unlike chat systems, they maintain state across interactions and translate intermediate outputs into concrete actions. This creates a distinct safety challenge in that harmful behavior may emerge through sequences of individually plausible steps, including intermediate actions that appear locally acceptable but collectively lead to unauthorized actions. We present AgentHazard, a benchmark for evaluating harmful behavior in computer-use agents. AgentHazard contains 2,653 instances spanning diverse risk categories and attack strategies. Each instance pairs a harmful objective with a sequence of operational steps that are locally legitimate but jointly induce unsafe behavior. The benchmark evaluates whether agents can recognize and interrupt harm arising from accumulated context, repeated tool use, intermediate actions, and dependencies across steps. We evaluate AgentHazard on Claude Code, OpenClaw, and IFlow using mostly open or openly deployable models from the Qwen3, Kimi, GLM, and DeepSeek families. Our experimental results indicate that current systems remain highly vulnerable. In particular, when powered by Qwen3-Coder, Claude Code exhibits an attack success rate of 73.63\%, suggesting that model alignment alone does not reliably guarantee the safety of autonomous agents.

Persuasion is a powerful capability of large language models (LLMs) that both enables beneficial applications (e.g. helping people quit smoking) and raises significant risks (e.g. large-scale, targeted political manipulation). Prior work has found models possess a significant and growing persuasive capability, measured by belief changes in simulated or real users. However, these benchmarks overlook a crucial risk factor: the propensity of a model to attempt to persuade in harmful contexts. Understanding whether a model will blindly ``follow orders'' to persuade on harmful topics (e.g. glorifying joining a terrorist group) is key to understanding the efficacy of safety guardrails. Moreover, understanding if and when a model will engage in persuasive behavior in pursuit of some goal is essential to understanding the risks from agentic AI systems. We propose the Attempt to Persuade Eval (APE) benchmark, that shifts the focus from persuasion success to persuasion attempts, operationalized as a model's willingness to generate content aimed at shaping beliefs or behavior. Our evaluation framework probes frontier LLMs using a multi-turn conversational setup between simulated persuader and persuadee agents. APE explores a diverse spectrum of topics including conspiracies, controversial issues, and non-controversially harmful content. We introduce an automated evaluator model to identify willingness to persuade and measure the frequency and context of persuasive attempts. We find that many open and closed-weight models are frequently willing to attempt persuasion on harmful topics and that jailbreaking can increase willingness to engage in such behavior. Our results highlight gaps in current safety guardrails and underscore the importance of evaluating willingness to persuade as a key dimension of LLM risk. APE is available at github.com/AlignmentResearch/AttemptPersuadeEval

Although large language models (LLMs) are typically aligned, they remain vulnerable to jailbreaking through either carefully crafted prompts in natural language or, interestingly, gibberish adversarial suffixes. However, gibberish tokens have received relatively less attention despite their success in attacking aligned LLMs. Recent work, AmpleGCG~liao2024amplegcg, demonstrates that a generative model can quickly produce numerous customizable gibberish adversarial suffixes for any harmful query, exposing a range of alignment gaps in out-of-distribution (OOD) language spaces. To bring more attention to this area, we introduce AmpleGCG-Plus, an enhanced version that achieves better performance in fewer attempts. Through a series of exploratory experiments, we identify several training strategies to improve the learning of gibberish suffixes. Our results, verified under a strict evaluation setting, show that it outperforms AmpleGCG on both open-weight and closed-source models, achieving increases in attack success rate (ASR) of up to 17\% in the white-box setting against Llama-2-7B-chat, and more than tripling ASR in the black-box setting against GPT-4. Notably, AmpleGCG-Plus jailbreaks the newer GPT-4o series of models at similar rates to GPT-4, and, uncovers vulnerabilities against the recently proposed circuit breakers defense. We publicly release AmpleGCG-Plus along with our collected training datasets.

Accurately evaluating adversarial robustness is a longstanding challenge. A flawed attack design can inflate robustness estimates, making deployment risk assessment and defense comparison unreliable. Historically, standardized attacks such as AutoAttack have largely resolved this for image classifiers, providing a reliable evaluation baseline for systematic comparison across defenses. However, no equivalent exists for LLM jailbreak evaluation yet, where designing such an attack is considerably more difficult. A reliable attack must, among other things, be black-box compatible, applicable to arbitrary defense pipelines, and efficient, which no existing method jointly satisfies. We introduce Indirect Harm Optimization (IHO), a masked diffusion language model attacker trained via iterative preference optimization against a harmfulness judge, requiring only black-box access to the target. The same method can be used without modification as a strong adaptive attack on individual behaviors, or as an efficient amortized policy that transfers to held-out behaviors and unseen target models without fine-tuning. Even against layered defenses, such as a Circuit Breaker-trained model combined with an auxiliary detector, IHO improves attack success considerably over state-of-the-art approaches, without any defense-specific adaptation. Our results position IHO as a practical step toward the kind of standardized jailbreak evaluation that has improved reliability in the past. Code and models are available on GitHub and Hugging Face.

As the application of generative adversarial networks (GANs) expands, it becomes increasingly critical to develop a unified approach that improves performance across various generative tasks. One effective strategy that applies to any machine learning task is identifying harmful instances, whose removal improves the performance. While previous studies have successfully estimated these harmful training instances in supervised settings, their approaches are not easily applicable to GANs. The challenge lies in two requirements of the previous approaches that do not apply to GANs. First, previous approaches require that the absence of a training instance directly affects the parameters. However, in the training for GANs, the instances do not directly affect the generator's parameters since they are only fed into the discriminator. Second, previous approaches assume that the change in loss directly quantifies the harmfulness of the instance to a model's performance, while common types of GAN losses do not always reflect the generative performance. To overcome the first challenge, we propose influence estimation methods that use the Jacobian of the generator's gradient with respect to the discriminator's parameters (and vice versa). Such a Jacobian represents the indirect effect between two models: how removing an instance from the discriminator's training changes the generator's parameters. Second, we propose an instance evaluation scheme that measures the harmfulness of each training instance based on how a GAN evaluation metric (e.g., Inception score) is expected to change by the instance's removal. Furthermore, we demonstrate that removing the identified harmful instances significantly improves the generative performance on various GAN evaluation metrics.

With the rapid development of large language models (LLMs), they are not only used as general-purpose AI assistants but are also customized through further fine-tuning to meet the requirements of different applications. A pivotal factor in the success of current LLMs is the alignment process. Current alignment methods, such as supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), focus on training-time alignment and are often complex and cumbersome to implement. Therefore, we develop InferAligner, a novel inference-time alignment method that utilizes cross-model guidance for harmlessness alignment. InferAligner utilizes safety steering vectors extracted from safety-aligned model to modify the activations of the target model when responding to harmful inputs, thereby guiding the target model to provide harmless responses. Experimental results show that our method can be very effectively applied to domain-specific models in finance, medicine, and mathematics, as well as to multimodal large language models (MLLMs) such as LLaVA. It significantly diminishes the Attack Success Rate (ASR) of both harmful instructions and jailbreak attacks, while maintaining almost unchanged performance in downstream tasks.

Large Language Model (LLM) unlearning is often evaluated primarily by target suppression and clean retain utility. These metrics can miss a local failure mode: a model may perform well on standard retain prompts, yet degrade on benign reasoning when prompts lie near the forgotten domain. This phenomenon is termed forget-adjacent utility loss. A fundamental question is: where should forget representations go after they leave the removed behavior? Successful unlearning depends not only on whether forget-conditioned representations depart from the harmful computation, but also on the direction in which they are redirected afterward. We propose RA^2 (Retain-Anchored Attraction), a latent-space unlearning method that redirects forget-token states toward nearby retain-supported representations while preserving higher-layer behavior on retain data. On WMDP and MUSE, RA^2 yields higher forget-adjacent utility than strong unlearning baselines at similar broad-utility levels. These results show that the migration direction of forget representations materially affects benign behavior near the forget boundary. Our code is available at https://github.com/ZacLing/retain_anchored_attraction.

The success of the reward model in distinguishing between responses with subtle safety differences depends critically on the high-quality preference dataset, which should capture the fine-grained nuances of harmful and harmless responses. This motivates the need to develop a dataset involving preference margins, which accurately quantify how harmless one response is compared to another. In this paper, we take the first step to propose an effective and cost-efficient framework to promote the margin-enhanced preference dataset development. Our framework, Legend, Leverages representation engineering to annotate preference datasets. It constructs the specific direction within the LLM's embedding space that represents safety. By leveraging this safety direction, Legend can then leverage the semantic distances of paired responses along this direction to annotate margins automatically. We experimentally demonstrate our effectiveness in both reward modeling and harmless alignment for LLMs. Legend also stands out for its efficiency, requiring only the inference time rather than additional training. This efficiency allows for easier implementation and scalability, making Legend particularly valuable for practical applications in aligning LLMs with safe conversations.

With the recent surge in popularity of LLMs has come an ever-increasing need for LLM safety training. In this paper, we investigate the fragility of SOTA open-source LLMs under simple, optimization-free attacks we refer to as priming attacks, which are easy to execute and effectively bypass alignment from safety training. Our proposed attack improves the Attack Success Rate on Harmful Behaviors, as measured by Llama Guard, by up to 3.3times compared to baselines. Source code and data are available at https://github.com/uiuc-focal-lab/llm-priming-attacks.

Large Language Models (LLMs) in multi-agent systems (MAS) have shown promise for complex tasks, yet current training methods lack principled ways to connect system-level evaluation with agent- and message-level learning. We propose a theoretical framework that unifies cooperative game-theoretic attribution with process reward modeling to transform system evaluation to agent credit to response-level signals. Unlike prior approaches that rely only on attribution (Shapley) or step-level labels (PRM), our method produces local, signed, and credit-conserving signals. In success cases, Shapley-based credit assignment fairly allocates outcomes across agents and is refined into per-message rewards that promote cooperation while discouraging redundancy or sabotage; in failure cases, first-error localization yields repair-aware preferences that penalize harmful steps while rewarding corrective attempts. The resulting signals are bounded, cooperative, and directly compatible with reinforcement- or preference-based post-training, providing a unified and auditable pathway from global evaluation to local supervision in LLM multi-agent training. Our contribution is conceptual: we present a theoretical foundation and training signals, leaving empirical validation for future work.

Safety evaluations of large language models commonly summarize harmful-output behavior with attack success rate (ASR). Yet the same non-harmful outcome can arise for very different reasons. A model may recover a harmful task and refuse it, fail to recover the task, or respond to something else entirely. Distinguishing these cases becomes especially important under intent-obscuring prompts, where a low ASR does not reveal whether the evaluated task was actually engaged. To make this distinction explicit, we pair ASR with operative understanding rate (UR), which measures whether a response both identifies the evaluated task and treats it as the task to be answered. Across interfaces, this paired view reveals substantial variation hidden by ASR: similar ASR values can correspond to sharply different recovery rates. Controlled English reconstructions show that recovery consistently improves as compressed prompts become more explicit, whereas ASR does not follow the same pattern. A complementary contrast comes from FormalLogic, where high recovery can still coincide with frequent harmful assistance. Together, these results show that non-harmful outcomes are not equally informative about model safety, motivating the joint reporting of intent recovery and ASR in LLM safety evaluation. Code and experiment inputs are available at https://github.com/kevinjiang0121-cyber/IRIS.

Rapid advancements of large language models (LLMs) have enabled the processing, understanding, and generation of human-like text, with increasing integration into systems that touch our social sphere. Despite this success, these models can learn, perpetuate, and amplify harmful social biases. In this paper, we present a comprehensive survey of bias evaluation and mitigation techniques for LLMs. We first consolidate, formalize, and expand notions of social bias and fairness in natural language processing, defining distinct facets of harm and introducing several desiderata to operationalize fairness for LLMs. We then unify the literature by proposing three intuitive taxonomies, two for bias evaluation, namely metrics and datasets, and one for mitigation. Our first taxonomy of metrics for bias evaluation disambiguates the relationship between metrics and evaluation datasets, and organizes metrics by the different levels at which they operate in a model: embeddings, probabilities, and generated text. Our second taxonomy of datasets for bias evaluation categorizes datasets by their structure as counterfactual inputs or prompts, and identifies the targeted harms and social groups; we also release a consolidation of publicly-available datasets for improved access. Our third taxonomy of techniques for bias mitigation classifies methods by their intervention during pre-processing, in-training, intra-processing, and post-processing, with granular subcategories that elucidate research trends. Finally, we identify open problems and challenges for future work. Synthesizing a wide range of recent research, we aim to provide a clear guide of the existing literature that empowers researchers and practitioners to better understand and prevent the propagation of bias in LLMs.

The widespread adoption of large language models (LLMs) has raised concerns about their safety and reliability, particularly regarding their vulnerability to adversarial attacks. In this paper, we propose a novel perspective that attributes this vulnerability to reward misspecification during the alignment process. We introduce a metric ReGap to quantify the extent of reward misspecification and demonstrate its effectiveness and robustness in detecting harmful backdoor prompts. Building upon these insights, we present ReMiss, a system for automated red teaming that generates adversarial prompts against various target aligned LLMs. ReMiss achieves state-of-the-art attack success rates on the AdvBench benchmark while preserving the human readability of the generated prompts. Detailed analysis highlights the unique advantages brought by the proposed reward misspecification objective compared to previous methods.

We introduce a low-resource safety enhancement method for aligning large language models (LLMs) without the need for supervised fine-tuning (SFT) or reinforcement learning from human feedback (RLHF). Our main idea is to exploit knowledge distillation to extract the alignment information from existing well-aligned LLMs and integrate it into unaligned LLMs in a plug-and-play fashion. Methodology, we employ delta debugging to identify the critical components of knowledge necessary for effective distillation. On the harmful question dataset, our method significantly enhances the average defense success rate by approximately 14.41%, reaching as high as 51.39%, in 17 unaligned pre-trained LLMs, without compromising performance.

Large Language Model safety alignment predominantly operates on a binary assumption that requests are either safe or unsafe. This classification proves insufficient when models encounter ethical dilemmas, where the capacity to reason through moral trade-offs creates a distinct attack surface. We formalize this vulnerability through TRIAL, a multi-turn red-teaming methodology that embeds harmful requests within ethical framings. TRIAL achieves high attack success rates across most tested models by systematically exploiting the model's ethical reasoning capabilities to frame harmful actions as morally necessary compromises. Building on these insights, we introduce ERR (Ethical Reasoning Robustness), a defense framework that distinguishes between instrumental responses that enable harmful outcomes and explanatory responses that analyze ethical frameworks without endorsing harmful acts. ERR employs a Layer-Stratified Harm-Gated LoRA architecture, achieving robust defense against reasoning-based attacks while preserving model utility.

This technical report accompanies the LLM robustness leaderboard published by PRISM Eval for the Paris AI Action Summit. We introduce PRISM Eval Behavior Elicitation Tool (BET), an AI system performing automated red-teaming through Dynamic Adversarial Optimization that achieves 100% Attack Success Rate (ASR) against 37 of 41 state-of-the-art LLMs. Beyond binary success metrics, we propose a fine-grained robustness metric estimating the average number of attempts required to elicit harmful behaviors, revealing that attack difficulty varies by over 300-fold across models despite universal vulnerability. We introduce primitive-level vulnerability analysis to identify which jailbreaking techniques are most effective for specific hazard categories. Our collaborative evaluation with trusted third parties from the AI Safety Network demonstrates practical pathways for distributed robustness assessment across the community.

Omni-modal Large Language Models (OLLMs) greatly expand LLMs' multimodal capabilities but also introduce cross-modal safety risks. However, a systematic understanding of vulnerabilities in omni-modal interactions remains lacking. To bridge this gap, we establish a modality-semantics decoupling principle and construct the AdvBench-Omni dataset, which reveals a significant vulnerability in OLLMs. Mechanistic analysis uncovers a Mid-layer Dissolution phenomenon driven by refusal vector magnitude shrinkage, alongside the existence of a modal-invariant pure refusal direction. Inspired by these insights, we extract a golden refusal vector using Singular Value Decomposition and propose OmniSteer, which utilizes lightweight adapters to modulate intervention intensity adaptively. Extensive experiments show that our method not only increases the Refusal Success Rate against harmful inputs from 69.9% to 91.2%, but also effectively preserves the general capabilities across all modalities. Our code is available at: https://github.com/zhrli324/omni-safety-research.

Jailbreak attacks, which aim to cause LLMs to perform unrestricted behaviors, have become a critical and challenging direction in AI safety. Despite achieving the promising attack success rate using dictionary-based evaluation, existing jailbreak attack methods fail to output detailed contents to satisfy the harmful request, leading to poor performance on GPT-based evaluation. To this end, we propose a black-box jailbreak attack termed GeneShift, by using a genetic algorithm to optimize the scenario shifts. Firstly, we observe that the malicious queries perform optimally under different scenario shifts. Based on it, we develop a genetic algorithm to evolve and select the hybrid of scenario shifts. It guides our method to elicit detailed and actionable harmful responses while keeping the seemingly benign facade, improving stealthiness. Extensive experiments demonstrate the superiority of GeneShift. Notably, GeneShift increases the jailbreak success rate from 0% to 60% when direct prompting alone would fail.

Recent studies demonstrate that Large Language Models (LLMs) are vulnerable to attacks that generate harmful or sensitive outputs. As open-source LLMs are increasingly adopted in high-impact applications such as finance, law, and healthcare, systematically investigating their security risks is becoming increasingly important towards trustworthy LLM era. This paper comprehensively studies effective prompt injection attacks against 14 widely used open-source and three closed-source LLMs on five attack benchmarks. Moreover, existing evaluation metrics mostly only consider the attack success rate, overlooking uncertainty in model responses. Our proposed Attack Success Probability (ASP) additionally captures uncertain behaviors for evaluation, where the model may initially refuse a harmful request but subsequently provide harmful guidance or vice versa, reflecting inconsistency and ambiguity in attack feasibility. By systematically analyzing the effectiveness of prompt injection attacks, we propose a straightforward and effective hypnotism attack; results show that this attack causes aligned language models, including Stablelm2, Mistral, Openchat, and Vicuna, to generate objectionable behaviors, achieving around 90% ASP. They also indicate that ignore prefix attacks can break all 14 open-source LLMs, achieving over 60% ASP on a multi-categorical dataset. We find that moderately well-known LLMs exhibit higher vulnerability to prompt injection attacks, highlighting the need to raise public awareness and prioritize efficient mitigation strategies.

As large language models (LLMs) become increasingly prevalent and integrated into autonomous systems, ensuring their safety is imperative. Despite significant strides toward safety alignment, recent work GCG~zou2023universal proposes a discrete token optimization algorithm and selects the single suffix with the lowest loss to successfully jailbreak aligned LLMs. In this work, we first discuss the drawbacks of solely picking the suffix with the lowest loss during GCG optimization for jailbreaking and uncover the missed successful suffixes during the intermediate steps. Moreover, we utilize those successful suffixes as training data to learn a generative model, named AmpleGCG, which captures the distribution of adversarial suffixes given a harmful query and enables the rapid generation of hundreds of suffixes for any harmful queries in seconds. AmpleGCG achieves near 100\% attack success rate (ASR) on two aligned LLMs (Llama-2-7B-chat and Vicuna-7B), surpassing two strongest attack baselines. More interestingly, AmpleGCG also transfers seamlessly to attack different models, including closed-source LLMs, achieving a 99\% ASR on the latest GPT-3.5. To summarize, our work amplifies the impact of GCG by training a generative model of adversarial suffixes that is universal to any harmful queries and transferable from attacking open-source LLMs to closed-source LLMs. In addition, it can generate 200 adversarial suffixes for one harmful query in only 4 seconds, rendering it more challenging to defend.

Diffusion models (DMs) have achieved remarkable success in text-to-image generation, but they also pose safety risks, such as the potential generation of harmful content and copyright violations. The techniques of machine unlearning, also known as concept erasing, have been developed to address these risks. However, these techniques remain vulnerable to adversarial prompt attacks, which can prompt DMs post-unlearning to regenerate undesired images containing concepts (such as nudity) meant to be erased. This work aims to enhance the robustness of concept erasing by integrating the principle of adversarial training (AT) into machine unlearning, resulting in the robust unlearning framework referred to as AdvUnlearn. However, achieving this effectively and efficiently is highly nontrivial. First, we find that a straightforward implementation of AT compromises DMs' image generation quality post-unlearning. To address this, we develop a utility-retaining regularization on an additional retain set, optimizing the trade-off between concept erasure robustness and model utility in AdvUnlearn. Moreover, we identify the text encoder as a more suitable module for robustification compared to UNet, ensuring unlearning effectiveness. And the acquired text encoder can serve as a plug-and-play robust unlearner for various DM types. Empirically, we perform extensive experiments to demonstrate the robustness advantage of AdvUnlearn across various DM unlearning scenarios, including the erasure of nudity, objects, and style concepts. In addition to robustness, AdvUnlearn also achieves a balanced tradeoff with model utility. To our knowledge, this is the first work to systematically explore robust DM unlearning through AT, setting it apart from existing methods that overlook robustness in concept erasing. Codes are available at: https://github.com/OPTML-Group/AdvUnlearn

LLMs are known to be vulnerable to jailbreak attacks, even after safety alignment. An important observation is that, while different types of jailbreak attacks can generate significantly different queries, they mostly result in similar responses that are rooted in the same harmful knowledge (e.g., detailed steps to make a bomb). Therefore, we conjecture that directly unlearn the harmful knowledge in the LLM can be a more effective way to defend against jailbreak attacks than the mainstream supervised fine-tuning (SFT) based approaches. Our extensive experiments confirmed our insight and suggested surprising generalizability of our unlearning-based approach: using only 20 raw harmful questions without any jailbreak prompt during training, our solution reduced the Attack Success Rate (ASR) in Vicuna-7B on out-of-distribution (OOD) harmful questions wrapped with various complex jailbreak prompts from 82.6\% to 7.7\%. This significantly outperforms Llama2-7B-Chat, which is fine-tuned on about 0.1M safety alignment samples but still has an ASR of 21.9\% even under the help of an additional safety system prompt. Further analysis reveals that the generalization ability of our solution stems from the intrinsic relatedness among harmful responses across harmful questions (e.g., response patterns, shared steps and actions, and similarity among their learned representations in the LLM). Our code is available at https://github.com/thu-coai/SafeUnlearning.

Large Language Models (LLMs) like GPT-4, LLaMA, and Qwen have demonstrated remarkable success across a wide range of applications. However, these models remain inherently vulnerable to prompt injection attacks, which can bypass existing safety mechanisms, highlighting the urgent need for more robust attack detection methods and comprehensive evaluation benchmarks. To address these challenges, we introduce GenTel-Safe, a unified framework that includes a novel prompt injection attack detection method, GenTel-Shield, along with a comprehensive evaluation benchmark, GenTel-Bench, which compromises 84812 prompt injection attacks, spanning 3 major categories and 28 security scenarios. To prove the effectiveness of GenTel-Shield, we evaluate it together with vanilla safety guardrails against the GenTel-Bench dataset. Empirically, GenTel-Shield can achieve state-of-the-art attack detection success rates, which reveals the critical weakness of existing safeguarding techniques against harmful prompts. For reproducibility, we have made the code and benchmarking dataset available on the project page at https://gentellab.github.io/gentel-safe.github.io/.

Jailbreaking large-language models (LLMs) involves testing their robustness against adversarial prompts and evaluating their ability to withstand prompt attacks that could elicit unauthorized or malicious responses. In this paper, we present TurboFuzzLLM, a mutation-based fuzzing technique for efficiently finding a collection of effective jailbreaking templates that, when combined with harmful questions, can lead a target LLM to produce harmful responses through black-box access via user prompts. We describe the limitations of directly applying existing template-based attacking techniques in practice, and present functional and efficiency-focused upgrades we added to mutation-based fuzzing to generate effective jailbreaking templates automatically. TurboFuzzLLM achieves geq 95\% attack success rates (ASR) on public datasets for leading LLMs (including GPT-4o \& GPT-4 Turbo), shows impressive generalizability to unseen harmful questions, and helps in improving model defenses to prompt attacks. TurboFuzzLLM is available open source at https://github.com/amazon-science/TurboFuzzLLM.

Large audio language models (LALMs) are increasingly deployed in real-world applications, yet their safety alignment is still primarily evaluated on monolingual, text-based harmful prompts. This leaves their generalizability under multilingual and spoken settings, particularly code-switched speech, largely underexplored. To address this gap, we introduce SpeechJBB, an audio jailbreak dataset for benchmarking across multiple state-of-the-art LALMs. The extent of safety weaknesses is further probed by introducing an augmented setting where phonologically plausible pseudo-words are inserted around safety-critical terms to simulate localized obfuscation. Across models, code-switched harmful audio yields substantially high jailbreak success rates (JSR), with non-English monolingual and non-English code-switched pairs exhibiting the highest attack success. Pseudo-word insertion further reduces refusal rates, which demonstrates that natural-sounding obfuscation can effectively bypass safety policies.

Large Language Models (LLMs) have achieved remarkable success but remain highly susceptible to jailbreak attacks, in which adversarial prompts coerce models into generating harmful, unethical, or policy-violating outputs. Such attacks pose real-world risks, eroding safety, trust, and regulatory compliance in high-stakes applications. Although a variety of attack and defense methods have been proposed, existing evaluation practices are inadequate, often relying on narrow metrics like attack success rate that fail to capture the multidimensional nature of LLM security. In this paper, we present a systematic taxonomy of jailbreak attacks and defenses and introduce Security Cube, a unified, multi-dimensional framework for comprehensive evaluation of these techniques. We provide detailed comparison tables of existing attacks and defenses, highlighting key insights and open challenges across the literature. Leveraging Security Cube, we conduct benchmark studies on 13 representative attacks and 5 defenses, establishing a clear view of the current landscape encompassing jailbreak attacks, defenses, automated judges, and LLM vulnerabilities. Based on these evaluations, we distill critical findings, identify unresolved problems, and outline promising research directions for enhancing LLM robustness against jailbreak attacks. Our analysis aims to pave the way towards more robust, interpretable, and trustworthy LLM systems. Our code is available at Code.

Full-duplex speech models accept user speech while generating responses, creating an underexplored attack surface. We introduce DuplexJail, which delivers fixed, request-independent spoken prompts through the user audio channel. We compare fixed-delay interruption after the harmful request ends with refusal-triggered interruption following a cue in the model's streaming text. Across four open-source models and 720 harmful requests from AdvBench and HarmBench, fixed-delay interruption raises whole-response attack success rates on AdvBench to 40.3% for PersonaPlex and 48.7% for PersonaPlex-RL, increases of +33.8 and +39.3 percentage points. The refusal-triggered policy reaches 35.6% and 48.6%, respectively, with all trials scored regardless of whether an interruption occurs. Selected conditions also increase FLM-Audio's harmful-response rate, while BayLing-Duplex shows decreases. These findings identify spoken interruption as a jailbreak attack vector and motivate evaluating safety throughout ongoing full-duplex interaction.

We introduce Best-of-N (BoN) Jailbreaking, a simple black-box algorithm that jailbreaks frontier AI systems across modalities. BoN Jailbreaking works by repeatedly sampling variations of a prompt with a combination of augmentations - such as random shuffling or capitalization for textual prompts - until a harmful response is elicited. We find that BoN Jailbreaking achieves high attack success rates (ASRs) on closed-source language models, such as 89% on GPT-4o and 78% on Claude 3.5 Sonnet when sampling 10,000 augmented prompts. Further, it is similarly effective at circumventing state-of-the-art open-source defenses like circuit breakers. BoN also seamlessly extends to other modalities: it jailbreaks vision language models (VLMs) such as GPT-4o and audio language models (ALMs) like Gemini 1.5 Pro, using modality-specific augmentations. BoN reliably improves when we sample more augmented prompts. Across all modalities, ASR, as a function of the number of samples (N), empirically follows power-law-like behavior for many orders of magnitude. BoN Jailbreaking can also be composed with other black-box algorithms for even more effective attacks - combining BoN with an optimized prefix attack achieves up to a 35% increase in ASR. Overall, our work indicates that, despite their capability, language models are sensitive to seemingly innocuous changes to inputs, which attackers can exploit across modalities.

We investigate long-context vulnerabilities in Large Language Models (LLMs) through Many-Shot Jailbreaking (MSJ). Our experiments utilize context length of up to 128K tokens. Through comprehensive analysis with various many-shot attack settings with different instruction styles, shot density, topic, and format, we reveal that context length is the primary factor determining attack effectiveness. Critically, we find that successful attacks do not require carefully crafted harmful content. Even repetitive shots or random dummy text can circumvent model safety measures, suggesting fundamental limitations in long-context processing capabilities of LLMs. The safety behavior of well-aligned models becomes increasingly inconsistent with longer contexts. These findings highlight significant safety gaps in context expansion capabilities of LLMs, emphasizing the need for new safety mechanisms.

Agents are deployed in applications from trip planners and stock trading to IT incident triage. In most cases, LLM agents work autonomously with minimal rule-based safeguarding, leading to cost and safety issues from irreversible actions. Recent works resolve this either by using a safeguard agent to monitor behavior or evaluating logs post-hoc. The first adds cost and latency to every step; the second delivers its verdict after the run, when tokens are burned and damage is done. To overcome this, we propose OnTrack, a streaming monitoring mechanism that compares an agent's steps and dependencies against recorded successful runs to alert users or block the agent in about a millisecond per step. We study this problem in three regimes of decreasing access: full reference access (historical runs and tool schemas), intermediate access (only tool schemas), and no prior knowledge (only step logs as generated). Expectation of OnTrack's monitoring capabilities reduces as data access drops, ranging from plan violation detection to identifying loops, stalls, and repeated tool calls. Finally, we evaluate OnTrack using SWE-bench trajectories. Based on the first 8 steps, our method ranks failing trajectories below succeeding ones better than content similarity approaches (+0.057 AUROC). With an abort policy, we save about 18% of compute that would be burned on failing runs, where 83% of interrupted runs were actually heading to failure (5 out of 6 aborts were correct).

Refusal training is widely used to prevent LLMs from generating harmful, undesirable, or illegal outputs. We reveal a curious generalization gap in the current refusal training approaches: simply reformulating a harmful request in the past tense (e.g., "How to make a Molotov cocktail?" to "How did people make a Molotov cocktail?") is often sufficient to jailbreak many state-of-the-art LLMs. We systematically evaluate this method on Llama-3 8B, Claude-3.5 Sonnet, GPT-3.5 Turbo, Gemma-2 9B, Phi-3-Mini, GPT-4o mini, GPT-4o, and R2D2 models using GPT-3.5 Turbo as a reformulation model. For example, the success rate of this simple attack on GPT-4o increases from 1% using direct requests to 88% using 20 past tense reformulation attempts on harmful requests from JailbreakBench with GPT-4 as a jailbreak judge. Interestingly, we also find that reformulations in the future tense are less effective, suggesting that refusal guardrails tend to consider past historical questions more benign than hypothetical future questions. Moreover, our experiments on fine-tuning GPT-3.5 Turbo show that defending against past reformulations is feasible when past tense examples are explicitly included in the fine-tuning data. Overall, our findings highlight that the widely used alignment techniques -- such as SFT, RLHF, and adversarial training -- employed to align the studied models can be brittle and do not always generalize as intended. We provide code and jailbreak artifacts at https://github.com/tml-epfl/llm-past-tense.

Multimodal Language Models (MMLMs) typically undergo post-training alignment to prevent harmful content generation. However, these alignment stages focus primarily on the assistant role, leaving the user role unaligned, and stick to a fixed input prompt structure of special tokens, leaving the model vulnerable when inputs deviate from these expectations. We introduce Role-Modality Attacks (RMA), a novel class of adversarial attacks that exploit role confusion between the user and assistant and alter the position of the image token to elicit harmful outputs. Unlike existing attacks that modify query content, RMAs manipulate the input structure without altering the query itself. We systematically evaluate these attacks across multiple Vision Language Models (VLMs) on eight distinct settings, showing that they can be composed to create stronger adversarial prompts, as also evidenced by their increased projection in the negative refusal direction in the residual stream, a property observed in prior successful attacks. Finally, for mitigation, we propose an adversarial training approach that makes the model robust against input prompt perturbations. By training the model on a range of harmful and benign prompts all perturbed with different RMA settings, it loses its sensitivity to Role Confusion and Modality Manipulation attacks and is trained to only pay attention to the content of the query in the input prompt structure, effectively reducing Attack Success Rate (ASR) while preserving the model's general utility.

Large language models (LLMs) have recently experienced tremendous popularity and are widely used from casual conversations to AI-driven programming. However, despite their considerable success, LLMs are not entirely reliable and can give detailed guidance on how to conduct harmful or illegal activities. While safety measures can reduce the risk of such outputs, adversarial jailbreak attacks can still exploit LLMs to produce harmful content. These jailbreak templates are typically manually crafted, making large-scale testing challenging. In this paper, we introduce GPTFuzz, a novel black-box jailbreak fuzzing framework inspired by the AFL fuzzing framework. Instead of manual engineering, GPTFuzz automates the generation of jailbreak templates for red-teaming LLMs. At its core, GPTFuzz starts with human-written templates as initial seeds, then mutates them to produce new templates. We detail three key components of GPTFuzz: a seed selection strategy for balancing efficiency and variability, mutate operators for creating semantically equivalent or similar sentences, and a judgment model to assess the success of a jailbreak attack. We evaluate GPTFuzz against various commercial and open-source LLMs, including ChatGPT, LLaMa-2, and Vicuna, under diverse attack scenarios. Our results indicate that GPTFuzz consistently produces jailbreak templates with a high success rate, surpassing human-crafted templates. Remarkably, GPTFuzz achieves over 90% attack success rates against ChatGPT and Llama-2 models, even with suboptimal initial seed templates. We anticipate that GPTFuzz will be instrumental for researchers and practitioners in examining LLM robustness and will encourage further exploration into enhancing LLM safety.

It has been shown that Large Language Model (LLM) alignments can be circumvented by appending specially crafted attack suffixes with harmful queries to elicit harmful responses. To conduct attacks against private target models whose characterization is unknown, public models can be used as proxies to fashion the attack, with successful attacks being transferred from public proxies to private target models. The success rate of attack depends on how closely the proxy model approximates the private model. We hypothesize that for attacks to be transferrable, it is sufficient if the proxy can approximate the target model in the neighborhood of the harmful query. Therefore, in this paper, we propose Local Fine-Tuning (LoFT), i.e., fine-tuning proxy models on similar queries that lie in the lexico-semantic neighborhood of harmful queries to decrease the divergence between the proxy and target models. First, we demonstrate three approaches to prompt private target models to obtain similar queries given harmful queries. Next, we obtain data for local fine-tuning by eliciting responses from target models for the generated similar queries. Then, we optimize attack suffixes to generate attack prompts and evaluate the impact of our local fine-tuning on the attack's success rate. Experiments show that local fine-tuning of proxy models improves attack transferability and increases attack success rate by 39%, 7%, and 0.5% (absolute) on target models ChatGPT, GPT-4, and Claude respectively.

Safety-aligned language models refuse harmful requests through learned refusal behaviors encoded in their internal representations. Recent activation-based jailbreaking methods circumvent these safety mechanisms by applying orthogonal projections to remove refusal directions, but these approaches treat refusal as a one-dimensional phenomenon and ignore the rich distributional structure of model activations. We introduce a principled framework based on optimal transport theory that transforms the entire distribution of harmful activations to match harmless ones. By combining PCA with closed-form Gaussian optimal transport, we achieve efficient computation in high-dimensional representation spaces while preserving essential geometric structure. Across six models (Llama-2, Llama-3.1, Qwen-2.5; 7B-32B parameters), our method achieves up to 11% higher attack success rates than state-of-the-art baselines while maintaining comparable perplexity, demonstrating superior preservation of model capabilities. Critically, we discover that layer-selective intervention (applying optimal transport to 1-2 carefully chosen layers at approximately 40-60% network depth) substantially outperforms full-network interventions, revealing that refusal mechanisms may be localized rather than distributed. Our analysis provides new insights into the geometric structure of safety representations and suggests that current alignment methods may be vulnerable to distributional attacks beyond simple direction removal.

While recently Large Language Models (LLMs) have achieved remarkable successes, they are vulnerable to certain jailbreaking attacks that lead to generation of inappropriate or harmful content. Manual red-teaming requires finding adversarial prompts that cause such jailbreaking, e.g. by appending a suffix to a given instruction, which is inefficient and time-consuming. On the other hand, automatic adversarial prompt generation often leads to semantically meaningless attacks that can easily be detected by perplexity-based filters, may require gradient information from the TargetLLM, or do not scale well due to time-consuming discrete optimization processes over the token space. In this paper, we present a novel method that uses another LLM, called the AdvPrompter, to generate human-readable adversarial prompts in seconds, sim800times faster than existing optimization-based approaches. We train the AdvPrompter using a novel algorithm that does not require access to the gradients of the TargetLLM. This process alternates between two steps: (1) generating high-quality target adversarial suffixes by optimizing the AdvPrompter predictions, and (2) low-rank fine-tuning of the AdvPrompter with the generated adversarial suffixes. The trained AdvPrompter generates suffixes that veil the input instruction without changing its meaning, such that the TargetLLM is lured to give a harmful response. Experimental results on popular open source TargetLLMs show state-of-the-art results on the AdvBench dataset, that also transfer to closed-source black-box LLM APIs. Further, we demonstrate that by fine-tuning on a synthetic dataset generated by AdvPrompter, LLMs can be made more robust against jailbreaking attacks while maintaining performance, i.e. high MMLU scores.

This paper introduces FRACTURED-SORRY-Bench, a framework for evaluating the safety of Large Language Models (LLMs) against multi-turn conversational attacks. Building upon the SORRY-Bench dataset, we propose a simple yet effective method for generating adversarial prompts by breaking down harmful queries into seemingly innocuous sub-questions. Our approach achieves a maximum increase of +46.22\% in Attack Success Rates (ASRs) across GPT-4, GPT-4o, GPT-4o-mini, and GPT-3.5-Turbo models compared to baseline methods. We demonstrate that this technique poses a challenge to current LLM safety measures and highlights the need for more robust defenses against subtle, multi-turn attacks.

Large Vision-Language Models (LVLMs) signify a groundbreaking paradigm shift within the Artificial Intelligence (AI) community, extending beyond the capabilities of Large Language Models (LLMs) by assimilating additional modalities (e.g., images). Despite this advancement, the safety of LVLMs remains adequately underexplored, with a potential overreliance on the safety assurances purported by their underlying LLMs. In this paper, we propose FigStep, a straightforward yet effective black-box jailbreak algorithm against LVLMs. Instead of feeding textual harmful instructions directly, FigStep converts the prohibited content into images through typography to bypass the safety alignment. The experimental results indicate that FigStep can achieve an average attack success rate of 82.50% on six promising open-source LVLMs. Not merely to demonstrate the efficacy of FigStep, we conduct comprehensive ablation studies and analyze the distribution of the semantic embeddings to uncover that the reason behind the success of FigStep is the deficiency of safety alignment for visual embeddings. Moreover, we compare FigStep with five text-only jailbreaks and four image-based jailbreaks to demonstrate the superiority of FigStep, i.e., negligible attack costs and better attack performance. Above all, our work reveals that current LVLMs are vulnerable to jailbreak attacks, which highlights the necessity of novel cross-modality safety alignment techniques. Our code and datasets are available at https://github.com/ThuCCSLab/FigStep .

Automated LLM-as-a-Judge frameworks have become the de facto standard for scalable evaluation across natural language processing. For instance, in safety evaluation, these judges are relied upon to evaluate harmfulness in order to benchmark the robustness of safety against adversarial attacks. However, we show that existing validation protocols fail to account for substantial distribution shifts inherent to red-teaming: diverse victim models exhibit distinct generation styles, attacks distort output patterns, and semantic ambiguity varies significantly across jailbreak scenarios. Through a comprehensive audit using 6642 human-verified labels, we reveal that the unpredictable interaction of these shifts often causes judge performance to degrade to near random chance. This stands in stark contrast to the high human agreement reported in prior work. Crucially, we find that many attacks inflate their success rates by exploiting judge insufficiencies rather than eliciting genuinely harmful content. To enable more reliable evaluation, we propose ReliableBench, a benchmark of behaviors that remain more consistently judgeable, and JudgeStressTest, a dataset designed to expose judge failures. Data available at: https://github.com/SchwinnL/LLMJudgeReliability.

LLM agents are evolving rapidly, powered by code execution, tools, and the recently introduced agent skills feature. Skills allow users to extend LLM applications with specialized third-party code, knowledge, and instructions. Although this can extend agent capabilities to new domains, it creates an increasingly complex agent supply chain, offering new surfaces for prompt injection attacks. We identify skill-based prompt injection as a significant threat and introduce SkillInject, a benchmark evaluating the susceptibility of widely-used LLM agents to injections through skill files. SkillInject contains 202 injection-task pairs with attacks ranging from obviously malicious injections to subtle, context-dependent attacks hidden in otherwise legitimate instructions. We evaluate frontier LLMs on SkillInject, measuring both security in terms of harmful instruction avoidance and utility in terms of legitimate instruction compliance. Our results show that today's agents are highly vulnerable with up to 80% attack success rate with frontier models, often executing extremely harmful instructions including data exfiltration, destructive action, and ransomware-like behavior. They furthermore suggest that this problem will not be solved through model scaling or simple input filtering, but that robust agent security will require context-aware authorization frameworks. Our benchmark is available at https://www.skill-inject.com/.

Multimodal Large Language Models (MLLMs) are increasingly being deployed as automated content moderators. Within this landscape, we uncover a critical threat: Adversarial Smuggling Attacks. Unlike adversarial perturbations (for misclassification) and adversarial jailbreaks (for harmful output generation), adversarial smuggling exploits the Human-AI capability gap. It encodes harmful content into human-readable visual formats that remain AI-unreadable, thereby evading automated detection and enabling the dissemination of harmful content. We classify smuggling attacks into two pathways: (1) Perceptual Blindness, disrupting text recognition; and (2) Reasoning Blockade, inhibiting semantic understanding despite successful text recognition. To evaluate this threat, we constructed SmuggleBench, the first comprehensive benchmark comprising 1,700 adversarial smuggling attack instances. Evaluations on SmuggleBench reveal that both proprietary (e.g., GPT-5) and open-source (e.g., Qwen3-VL) state-of-the-art models are vulnerable to this threat, producing Attack Success Rates (ASR) exceeding 90%. By analyzing the vulnerability through the lenses of perception and reasoning, we identify three root causes: the limited capabilities of vision encoders, the robustness gap in OCR, and the scarcity of domain-specific adversarial examples. We conduct a preliminary exploration of mitigation strategies, investigating the potential of test-time scaling (via CoT) and adversarial training (via SFT) to mitigate this threat. Our code is publicly available at https://github.com/zhihengli-casia/smugglebench.

Tool-using LLM agents fail through trajectories rather than only final responses, as they may execute unsafe tool calls, follow injected instructions, comply with harmful requests, or over-refuse benign tasks despite producing a seemingly safe answer. Existing safety-alignment signals are largely response-level or off-policy, and often incur a safety-utility trade-off: improving agent safety comes at the cost of degraded task performance. Such sparse and single-objective rewards severely limit real-world usability. To bridge this gap, we propose FATE, an on-policy self-evolving framework that transforms verifier-scored failures into repair supervision without expert demonstrations. For each failure, the same policy proposes repair candidates, which are then re-scored by verifiers and filtered across security, utility, over-refusal control, and trajectory validity. This dense trajectory-level information is then used as a supervision signal for agent self-evolution. During this process, we further introduce Pareto-Front Policy Optimization (PFPO), combining supervised warmup with Pareto-aware policy optimization to preserve safety-utility trade-offs. Experiments on AgentDojo, AgentHarm, and ATBench show that FATE improves safety across different models and scales while preserving useful behavior. Compared with strong baselines, FATE reduces attack success rate by 33.5%, harmful compliance by 82.6%, and improves external trajectory-safety diagnosis by 6.5%. These results suggest that failed trajectories can provide structured repair supervision for safer self-evolving agents.

The predominant approach for training web navigation agents is to gather human demonstrations for a set of popular websites and hand-written tasks, but it is becoming clear that human data is an inefficient resource. We develop a pipeline to facilitate internet-scale training for agents without laborious human annotations. In the first stage, an LLM annotates 150k sites with agentic tasks. In the next stage, LLM agents complete tasks and produce trajectories. In the final stage, an LLM filters trajectories by judging their success. Language models are powerful data curation tools, identifying harmful content with an accuracy of 97%, judging successful trajectories with an accuracy of 82.6%, and producing effective data. We train agents based on Qwen 3 1.7B that are competitive with frontier LLMs as web agents, while being smaller and faster. Our top agent reaches a success rate of 56.9%, outperforming the data collection policy Qwen 3 235B, a 235 times larger Llama 4 Maverick, and reaching 94.7% of the performance of Gemini 2.5 Flash. We are releasing code, models and data at: https://data-for-agents.github.io.

Current LLM unlearning methods face a critical security vulnerability that undermines their fundamental purpose: while they appear to successfully remove sensitive or harmful knowledge, this ``forgotten" information remains precariously recoverable through relearning attacks. We identify that the root cause is that conventional methods optimizing the forgetting loss at individual data points will drive model parameters toward sharp minima in the loss landscape. In these unstable regions, even minimal parameter perturbations can drastically alter the model's behaviors. Consequently, relearning attacks exploit this vulnerability by using just a few fine-tuning samples to navigate the steep gradients surrounding these unstable regions, thereby rapidly recovering knowledge that was supposedly erased. This exposes a critical robustness gap between apparent unlearning and actual knowledge removal. To address this issue, we propose StableUN, a bi-level feedback-guided optimization framework that explicitly seeks more stable parameter regions via neighborhood-aware optimization. It integrates forgetting feedback, which uses adversarial perturbations to probe parameter neighborhoods, with remembering feedback to preserve model utility, aligning the two objectives through gradient projection. Experiments on WMDP and MUSE benchmarks demonstrate that our method is significantly more robust against both relearning and jailbreaking attacks while maintaining competitive utility performance.

Red-teaming has been a widely adopted way to evaluate the harmfulness of Large Language Models (LLMs). It aims to jailbreak a model's safety behavior to make it act as a helpful agent disregarding the harmfulness of the query. Existing methods are primarily based on input text-based red-teaming such as adversarial prompts, low-resource prompts, or contextualized prompts to condition the model in a way to bypass its safe behavior. Bypassing the guardrails uncovers hidden harmful information and biases in the model that are left untreated or newly introduced by its safety training. However, prompt-based attacks fail to provide such a diagnosis owing to their low attack success rate, and applicability to specific models. In this paper, we present a new perspective on LLM safety research i.e., parametric red-teaming through Unalignment. It simply (instruction) tunes the model parameters to break model guardrails that are not deeply rooted in the model's behavior. Unalignment using as few as 100 examples can significantly bypass commonly referred to as CHATGPT, to the point where it responds with an 88% success rate to harmful queries on two safety benchmark datasets. On open-source models such as VICUNA-7B and LLAMA-2-CHAT 7B AND 13B, it shows an attack success rate of more than 91%. On bias evaluations, Unalignment exposes inherent biases in safety-aligned models such as CHATGPT and LLAMA- 2-CHAT where the model's responses are strongly biased and opinionated 64% of the time.

Black-box finetuning is an emerging interface for adapting state-of-the-art language models to user needs. However, such access may also let malicious actors undermine model safety. To demonstrate the challenge of defending finetuning interfaces, we introduce covert malicious finetuning, a method to compromise model safety via finetuning while evading detection. Our method constructs a malicious dataset where every individual datapoint appears innocuous, but finetuning on the dataset teaches the model to respond to encoded harmful requests with encoded harmful responses. Applied to GPT-4, our method produces a finetuned model that acts on harmful instructions 99% of the time and avoids detection by defense mechanisms such as dataset inspection, safety evaluations, and input/output classifiers. Our findings question whether black-box finetuning access can be secured against sophisticated adversaries.

Large Language Models (LLMs) presents significant priority in text understanding and generation. However, LLMs suffer from the risk of generating harmful contents especially while being employed to applications. There are several black-box attack methods, such as Prompt Attack, which can change the behaviour of LLMs and induce LLMs to generate unexpected answers with harmful contents. Researchers are interested in Prompt Attack and Defense with LLMs, while there is no publicly available dataset with high successful attacking rate to evaluate the abilities of defending prompt attack. In this paper, we introduce a pipeline to construct high-quality prompt attack samples, along with a Chinese prompt attack dataset called CPAD. Our prompts aim to induce LLMs to generate unexpected outputs with several carefully designed prompt attack templates and widely concerned attacking contents. Different from previous datasets involving safety estimation, we construct the prompts considering three dimensions: contents, attacking methods and goals. Especially, the attacking goals indicate the behaviour expected after successfully attacking the LLMs, thus the responses can be easily evaluated and analysed. We run several popular Chinese LLMs on our dataset, and the results show that our prompts are significantly harmful to LLMs, with around 70% attack success rate to GPT-3.5. CPAD is publicly available at https://github.com/liuchengyuan123/CPAD.

This paper proposes a simple yet effective jailbreak attack named FlipAttack against black-box LLMs. First, from the autoregressive nature, we reveal that LLMs tend to understand the text from left to right and find that they struggle to comprehend the text when noise is added to the left side. Motivated by these insights, we propose to disguise the harmful prompt by constructing left-side noise merely based on the prompt itself, then generalize this idea to 4 flipping modes. Second, we verify the strong ability of LLMs to perform the text-flipping task, and then develop 4 variants to guide LLMs to denoise, understand, and execute harmful behaviors accurately. These designs keep FlipAttack universal, stealthy, and simple, allowing it to jailbreak black-box LLMs within only 1 query. Experiments on 8 LLMs demonstrate the superiority of FlipAttack. Remarkably, it achieves sim98\% attack success rate on GPT-4o, and sim98\% bypass rate against 5 guardrail models on average. The codes are available at GitHubhttps://github.com/yueliu1999/FlipAttack.

Along with the remarkable successes of Language language models, recent research also started to explore the security threats of LLMs, including jailbreaking attacks. Attackers carefully craft jailbreaking prompts such that a target LLM will respond to the harmful question. Existing jailbreaking attacks require either human experts or leveraging complicated algorithms to craft jailbreaking prompts. In this paper, we introduce BOOST, a simple attack that leverages only the eos tokens. We demonstrate that rather than constructing complicated jailbreaking prompts, the attacker can simply append a few eos tokens to the end of a harmful question. It will bypass the safety alignment of LLMs and lead to successful jailbreaking attacks. We further apply BOOST to four representative jailbreak methods and show that the attack success rates of these methods can be significantly enhanced by simply adding eos tokens to the prompt. To understand this simple but novel phenomenon, we conduct empirical analyses. Our analysis reveals that adding eos tokens makes the target LLM believe the input is much less harmful, and eos tokens have low attention values and do not affect LLM's understanding of the harmful questions, leading the model to actually respond to the questions. Our findings uncover how fragile an LLM is against jailbreak attacks, motivating the development of strong safety alignment approaches.

As large language models (LLMs) are increasingly deployed in diverse applications, including chatbot assistants and code generation, aligning their behavior with safety and ethical standards has become paramount. However, jailbreak attacks, which exploit vulnerabilities to elicit unintended or harmful outputs, threaten LLMs' safety significantly. In this paper, we introduce Layer-AdvPatcher, a novel methodology designed to defend against jailbreak attacks by utilizing an unlearning strategy to patch specific layers within LLMs through self-augmented datasets. Our insight is that certain layer(s), tend to produce affirmative tokens when faced with harmful prompts. By identifying these layers and adversarially exposing them to generate more harmful data, one can understand their inherent and diverse vulnerabilities to attacks. With these exposures, we then "unlearn" these issues, reducing the impact of affirmative tokens and hence minimizing jailbreak risks while keeping the model's responses to safe queries intact. We conduct extensive experiments on two models, four benchmark datasets, and multiple state-of-the-art jailbreak benchmarks to demonstrate the efficacy of our approach. Results indicate that our framework reduces the harmfulness and attack success rate of jailbreak attacks without compromising utility for benign queries compared to recent defense methods.

Large Vision-Language Models (LVLMs) have made significant strides in multimodal comprehension, thanks to extensive pre-training and fine-tuning on large-scale visual datasets. However, despite their robust textual safety mechanisms, they remain vulnerable to harmful visual inputs. Existing safeguards-typically relying on pre-filtering or fine-tuning-incur high costs and diminish overall utility. To address this critical vulnerability, we introduce SafeCLIP, a lightweight method that leverages LVLMs inherent multimodal alignment for zero-shot toxic image detection. By projecting CLIPs discarded CLS token into its text space and matching it with toxic descriptors, SafeCLIP detects harmful content without any architectural changes-adding minimal latency and enabling dynamic safety corrections during inference and fine-tuning.Experiments show that SafeCLIP achieves a 66.9% defense success rate with only 3.2% false positive rate and 7.2% overhead. In contrast, state-of-the-art methods achieve 52.9% success but have a 10.7% false positive rate and 210% overhead. Our work demonstrates that leveraging inherent multimodal alignment can yield efficient, low-cost LVLM safety. Code is available at anonymous.4open.science/r/safeclip-2C01.

Recent advances in long-context language models (LMs) have enabled million-token inputs, expanding their capabilities across complex tasks like computer-use agents. Yet, the safety implications of these extended contexts remain unclear. To bridge this gap, we introduce NINJA (short for Needle-in-haystack jailbreak attack), a method that jailbreaks aligned LMs by appending benign, model-generated content to harmful user goals. Critical to our method is the observation that the position of harmful goals play an important role in safety. Experiments on standard safety benchmark, HarmBench, show that NINJA significantly increases attack success rates across state-of-the-art open and proprietary models, including LLaMA, Qwen, Mistral, and Gemini. Unlike prior jailbreaking methods, our approach is low-resource, transferable, and less detectable. Moreover, we show that NINJA is compute-optimal -- under a fixed compute budget, increasing context length can outperform increasing the number of trials in best-of-N jailbreak. These findings reveal that even benign long contexts -- when crafted with careful goal positioning -- introduce fundamental vulnerabilities in modern LMs.

Computer-use agents (CUAs) can now autonomously complete complex tasks in real digital environments, but when misled, they can also be used to automate harmful actions programmatically. Existing safety evaluations largely target explicit threats such as misuse and prompt injection, but overlook a subtle yet critical setting where user instructions are entirely benign and harm arises from the task context or execution outcome. We introduce OS-BLIND, a benchmark that evaluates CUAs under unintended attack conditions, comprising 300 human-crafted tasks across 12 categories, 8 applications, and 2 threat clusters: environment-embedded threats and agent-initiated harms. Our evaluation on frontier models and agentic frameworks reveals that most CUAs exceed 90% attack success rate (ASR), and even the safety-aligned Claude 4.5 Sonnet reaches 73.0% ASR. More interestingly, this vulnerability becomes even more severe, with ASR rising from 73.0% to 92.7% when Claude 4.5 Sonnet is deployed in multi-agent systems. Our analysis further shows that existing safety defenses provide limited protection when user instructions are benign. Safety alignment primarily activates within the first few steps and rarely re-engages during subsequent execution. In multi-agent systems, decomposed subtasks obscure the harmful intent from the model, causing safety-aligned models to fail. We will release our OS-BLIND to encourage the broader research community to further investigate and address these safety challenges.

Direct Preference Optimisation (DPO) is widely used for safety alignment in large language models. However, prior work shows it is brittle and exhibits poor out-of-distribution (OOD) generalisation. In this paper, we investigate whether Curriculum Learning can improve the robustness of DPO-based safety alignment. We propose Staged-Competence, a curriculum-based framework that organises preference data by difficulty, employs competence-based sampling, and progressively updates the reference model during training. Averaged across three model families, Staged-Competence reduces OOD harmful response rates by 16% and jailbreak attack success rates by 20%, while preserving general capabilities with near-zero over-refusal. We further show that Staged-Competence (1) matches baseline safety with only 75% of the training data and (2) yields better separation between safe and unsafe responses. Staged-Competence is agnostic to the policy optimisation loss and can extend to other DPO variants and alignment domains. Our code and data are available at https://github.com/Sandeep5500/curriculum-learning-for-safety.

Federated learning offers a framework of training a machine learning model in a distributed fashion while preserving privacy of the participants. As the server cannot govern the clients' actions, nefarious clients may attack the global model by sending malicious local gradients. In the meantime, there could also be unreliable clients who are benign but each has a portion of low-quality training data (e.g., blur or low-resolution images), thus may appearing similar as malicious clients. Therefore, a defense mechanism will need to perform a three-fold differentiation which is much more challenging than the conventional (two-fold) case. This paper introduces MUD-HoG, a novel defense algorithm that addresses this challenge in federated learning using long-short history of gradients, and treats the detected malicious and unreliable clients differently. Not only this, but we can also distinguish between targeted and untargeted attacks among malicious clients, unlike most prior works which only consider one type of the attacks. Specifically, we take into account sign-flipping, additive-noise, label-flipping, and multi-label-flipping attacks, under a non-IID setting. We evaluate MUD-HoG with six state-of-the-art methods on two datasets. The results show that MUD-HoG outperforms all of them in terms of accuracy as well as precision and recall, in the presence of a mixture of multiple (four) types of attackers as well as unreliable clients. Moreover, unlike most prior works which can only tolerate a low population of harmful users, MUD-HoG can work with and successfully detect a wide range of malicious and unreliable clients - up to 47.5% and 10%, respectively, of the total population. Our code is open-sourced at https://github.com/LabSAINT/MUD-HoG_Federated_Learning.

The study of representation universality in AI models reveals growing convergence across domains, modalities, and architectures. However, the practical applications of representation universality remain largely unexplored. We bridge this gap by demonstrating that safety interventions can be transferred between models through learned mappings of their shared activation spaces. We demonstrate this approach on two well-established AI safety tasks: backdoor removal and refusal of harmful prompts, showing successful transfer of steering vectors that alter the models' outputs in a predictable way. Additionally, we propose a new task, corrupted capabilities, where models are fine-tuned to embed knowledge tied to a backdoor. This tests their ability to separate useful skills from backdoors, reflecting real-world challenges. Extensive experiments across Llama, Qwen and Gemma model families show that our method enables using smaller models to efficiently align larger ones. Furthermore, we demonstrate that autoencoder mappings between base and fine-tuned models can serve as reliable ``lightweight safety switches", allowing dynamic toggling between model behaviors.

With the widespread use of multi-modal Large Language models (MLLMs), safety issues have become a growing concern. Multi-turn dialogues, which are more common in everyday interactions, pose a greater risk than single prompts; however, existing benchmarks do not adequately consider this situation. To encourage the community to focus on the safety issues of these models in multi-turn dialogues, we introduce SafeMT, a benchmark that features dialogues of varying lengths generated from harmful queries accompanied by images. This benchmark consists of 10,000 samples in total, encompassing 17 different scenarios and four jailbreak methods. Additionally, we propose Safety Index (SI) to evaluate the general safety of MLLMs during conversations. We assess the safety of 17 models using this benchmark and discover that the risk of successful attacks on these models increases as the number of turns in harmful dialogues rises. This observation indicates that the safety mechanisms of these models are inadequate for recognizing the hazard in dialogue interactions. We propose a dialogue safety moderator capable of detecting malicious intent concealed within conversations and providing MLLMs with relevant safety policies. Experimental results from several open-source models indicate that this moderator is more effective in reducing multi-turn ASR compared to existed guard models.

Despite substantial efforts in safety alignment, recent research indicates that Large Language Models (LLMs) remain highly susceptible to jailbreak attacks. Among these attacks, finetuning-based ones that compromise LLMs' safety alignment via fine-tuning stand out due to its stable jailbreak performance. In particular, a recent study indicates that fine-tuning with as few as 10 harmful question-answer (QA) pairs can lead to successful jailbreaking across various harmful questions. However, such malicious fine-tuning attacks are readily detectable and hence thwarted by moderation models. In this paper, we demonstrate that LLMs can be jailbroken by fine-tuning with only 10 benign QA pairs; our attack exploits the increased sensitivity of LLMs to fine-tuning data after being overfitted. Specifically, our fine-tuning process starts with overfitting an LLM via fine-tuning with benign QA pairs involving identical refusal answers. Further fine-tuning is then performed with standard benign answers, causing the overfitted LLM to forget the refusal attitude and thus provide compliant answers regardless of the harmfulness of a question. We implement our attack on the ten LLMs and compare it with five existing baselines. Experiments demonstrate that our method achieves significant advantages in both attack effectiveness and attack stealth. Our findings expose previously unreported security vulnerabilities in current LLMs and provide a new perspective on understanding how LLMs' security is compromised, even with benign fine-tuning. Our code is available at https://github.com/ZHIXINXIE/tenBenign.

Cooperative language-model agents can improve without weight updates by revising reusable skills, but only if a shared team reward is traced to the responsible agent and skill. Existing methods credit the skills present in each trajectory, yet presence is not responsibility: a harmful skill active in successful runs still earns positive credit. We present GRSE, Group-Relative Skill Evolution, which assigns behavioral credit by comparison: inspired by GRPO, it executes each question several times under the same skill libraries and contrasts the executions to credit agent--skill pairs, with structured diagnosis of failures. On HotpotQA and LoCoMo with two open-weight actors, GRSE outperforms MASkills, the closest prior method, on HotpotQA and every LoCoMo category except open-domain, with the largest gains on multi-hop questions (+6.7 and +3.2 F1). Its gains rest on better-grounded retrieval rather than answers given without evidence, and its learned skills transfer between the multi-hop tasks. GRSE is robust to corrupted libraries: starting from planted harmful skills, it credits every planted skill with the correct sign and recovers the score of a clean library (+6.3 points), whereas MASkills misattributes two and falls below its starting point.

Batch Normalization (BN) has become an out-of-box technique to improve deep network training. However, its effectiveness is limited for micro-batch training, i.e., each GPU typically has only 1-2 images for training, which is inevitable for many computer vision tasks, e.g., object detection and semantic segmentation, constrained by memory consumption. To address this issue, we propose Weight Standardization (WS) and Batch-Channel Normalization (BCN) to bring two success factors of BN into micro-batch training: 1) the smoothing effects on the loss landscape and 2) the ability to avoid harmful elimination singularities along the training trajectory. WS standardizes the weights in convolutional layers to smooth the loss landscape by reducing the Lipschitz constants of the loss and the gradients; BCN combines batch and channel normalizations and leverages estimated statistics of the activations in convolutional layers to keep networks away from elimination singularities. We validate WS and BCN on comprehensive computer vision tasks, including image classification, object detection, instance segmentation, video recognition and semantic segmentation. All experimental results consistently show that WS and BCN improve micro-batch training significantly. Moreover, using WS and BCN with micro-batch training is even able to match or outperform the performances of BN with large-batch training.

Integrated Speech and Large Language Models (SLMs) that can follow speech instructions and generate relevant text responses have gained popularity lately. However, the safety and robustness of these models remains largely unclear. In this work, we investigate the potential vulnerabilities of such instruction-following speech-language models to adversarial attacks and jailbreaking. Specifically, we design algorithms that can generate adversarial examples to jailbreak SLMs in both white-box and black-box attack settings without human involvement. Additionally, we propose countermeasures to thwart such jailbreaking attacks. Our models, trained on dialog data with speech instructions, achieve state-of-the-art performance on spoken question-answering task, scoring over 80% on both safety and helpfulness metrics. Despite safety guardrails, experiments on jailbreaking demonstrate the vulnerability of SLMs to adversarial perturbations and transfer attacks, with average attack success rates of 90% and 10% respectively when evaluated on a dataset of carefully designed harmful questions spanning 12 different toxic categories. However, we demonstrate that our proposed countermeasures reduce the attack success significantly.

Security concerns related to Large Language Models (LLMs) have been extensively explored, yet the safety implications for Multimodal Large Language Models (MLLMs), particularly in medical contexts (MedMLLMs), remain insufficiently studied. This paper delves into the underexplored security vulnerabilities of MedMLLMs, especially when deployed in clinical environments where the accuracy and relevance of question-and-answer interactions are critically tested against complex medical challenges. By combining existing clinical medical data with atypical natural phenomena, we redefine two types of attacks: mismatched malicious attack (2M-attack) and optimized mismatched malicious attack (O2M-attack). Using our own constructed voluminous 3MAD dataset, which covers a wide range of medical image modalities and harmful medical scenarios, we conduct a comprehensive analysis and propose the MCM optimization method, which significantly enhances the attack success rate on MedMLLMs. Evaluations with this dataset and novel attack methods, including white-box attacks on LLaVA-Med and transfer attacks on four other state-of-the-art models, indicate that even MedMLLMs designed with enhanced security features are vulnerable to security breaches. Our work underscores the urgent need for a concerted effort to implement robust security measures and enhance the safety and efficacy of open-source MedMLLMs, particularly given the potential severity of jailbreak attacks and other malicious or clinically significant exploits in medical settings. For further research and replication, anonymous access to our code is available at https://github.com/dirtycomputer/O2M_attack. Warning: Medical large model jailbreaking may generate content that includes unverified diagnoses and treatment recommendations. Always consult professional medical advice.

Recently, the DeepSeek series of models, leveraging their exceptional reasoning capabilities and open-source strategy, is reshaping the global AI landscape. Despite these advantages, they exhibit significant safety deficiencies. Research conducted by Robust Intelligence, a subsidiary of Cisco, in collaboration with the University of Pennsylvania, revealed that DeepSeek-R1 has a 100\% attack success rate when processing harmful prompts. Additionally, multiple safety companies and research institutions have confirmed critical safety vulnerabilities in this model. As models demonstrating robust performance in Chinese and English, DeepSeek models require equally crucial safety assessments in both language contexts. However, current research has predominantly focused on safety evaluations in English environments, leaving a gap in comprehensive assessments of their safety performance in Chinese contexts. In response to this gap, this study introduces CHiSafetyBench, a Chinese-specific safety evaluation benchmark. This benchmark systematically evaluates the safety of DeepSeek-R1 and DeepSeek-V3 in Chinese contexts, revealing their performance across safety categories. The experimental results quantify the deficiencies of these two models in Chinese contexts, providing key insights for subsequent improvements. It should be noted that, despite our efforts to establish a comprehensive, objective, and authoritative evaluation benchmark, the selection of test samples, characteristics of data distribution, and the setting of evaluation criteria may inevitably introduce certain biases into the evaluation results. We will continuously optimize the evaluation benchmark and periodically update this report to provide more comprehensive and accurate assessment outcomes. Please refer to the latest version of the paper for the most recent evaluation results and conclusions.

Ensuring AI safety is crucial as large language models become increasingly integrated into real-world applications. A key challenge is jailbreak, where adversarial prompts bypass built-in safeguards to elicit harmful disallowed outputs. Inspired by psychological foot-in-the-door principles, we introduce FITD,a novel multi-turn jailbreak method that leverages the phenomenon where minor initial commitments lower resistance to more significant or more unethical transgressions. Our approach progressively escalates the malicious intent of user queries through intermediate bridge prompts and aligns the model's response by itself to induce toxic responses. Extensive experimental results on two jailbreak benchmarks demonstrate that FITD achieves an average attack success rate of 94% across seven widely used models, outperforming existing state-of-the-art methods. Additionally, we provide an in-depth analysis of LLM self-corruption, highlighting vulnerabilities in current alignment strategies and emphasizing the risks inherent in multi-turn interactions. The code is available at https://github.com/Jinxiaolong1129/Foot-in-the-door-Jailbreak.

Multimodal large language models (MLLMs) are widely used in vision-language reasoning tasks. However, their vulnerability to adversarial prompts remains a serious concern, as safety mechanisms often fail to prevent the generation of harmful outputs. Although recent jailbreak strategies report high success rates, many responses classified as "successful" are actually benign, vague, or unrelated to the intended malicious goal. This mismatch suggests that current evaluation standards may overestimate the effectiveness of such attacks. To address this issue, we introduce a four-axis evaluation framework that considers input on-topicness, input out-of-distribution (OOD) intensity, output harmfulness, and output refusal rate. This framework identifies truly effective jailbreaks. In a substantial empirical study, we reveal a structural trade-off: highly on-topic prompts are frequently blocked by safety filters, whereas those that are too OOD often evade detection but fail to produce harmful content. However, prompts that balance relevance and novelty are more likely to evade filters and trigger dangerous output. Building on this insight, we develop a recursive rewriting strategy called Balanced Structural Decomposition (BSD). The approach restructures malicious prompts into semantically aligned sub-tasks, while introducing subtle OOD signals and visual cues that make the inputs harder to detect. BSD was tested across 13 commercial and open-source MLLMs, where it consistently led to higher attack success rates, more harmful outputs, and fewer refusals. Compared to previous methods, it improves success rates by 67% and harmfulness by 21%, revealing a previously underappreciated weakness in current multimodal safety systems.

Modern language models often rely on Reinforcement Learning from Human Feedback (RLHF) to encourage safe behaviors. However, they remain vulnerable to adversarial attacks due to three key limitations: (1) the inefficiency and high cost of human annotation, (2) the vast diversity of potential adversarial attacks, and (3) the risk of feedback bias and reward hacking. To address these challenges, we introduce Adversarial Preference Learning (APL), an iterative adversarial training method incorporating three key innovations. First, a direct harmfulness metric based on the model's intrinsic preference probabilities, eliminating reliance on external assessment. Second, a conditional generative attacker that synthesizes input-specific adversarial variations. Third, an iterative framework with automated closed-loop feedback, enabling continuous adaptation through vulnerability discovery and mitigation. Experiments on Mistral-7B-Instruct-v0.3 demonstrate that APL significantly enhances robustness, achieving 83.33% harmlessness win rate over the base model (evaluated by GPT-4o), reducing harmful outputs from 5.88% to 0.43% (measured by LLaMA-Guard), and lowering attack success rate by up to 65% according to HarmBench. Notably, APL maintains competitive utility, with an MT-Bench score of 6.59 (comparable to the baseline 6.78) and an LC-WinRate of 46.52% against the base model.

Vision-language models (VLMs) achieve remarkable performance but remain vulnerable to adversarial attacks. Entropy, a measure of model uncertainty, is strongly correlated with the reliability of VLM. Prior entropy-based attacks maximize uncertainty at all decoding steps, implicitly assuming that every token contributes equally to generation instability. We show instead that a small fraction (about 20%) of high-entropy tokens, i.e., critical decision points in autoregressive generation, disproportionately governs output trajectories. By concentrating adversarial perturbations on these positions, we achieve semantic degradation comparable to global methods while using substantially smaller budgets. More importantly, across multiple representative VLMs, such selective attacks convert 35-49% of benign outputs into harmful ones, exposing a more critical safety risk. Remarkably, these vulnerable high-entropy forks recur across architecturally diverse VLMs, enabling feasible transferability (17-26% harmful rates on unseen targets). Motivated by these findings, we propose Entropy-bank Guided Adversarial attacks (EGA), which achieves competitive attack success rates (93-95%) alongside high harmful conversion, thereby revealing new weaknesses in current VLM safety mechanisms.

When a phone-use agent avoids harm, does that show safety, or simply inability to act? Existing evaluations often cannot tell. A harmful outcome may be avoided because the agent recognized the risk and chose the safe action, or because it failed to understand the screen or execute any relevant action at all. These cases have different causes and call for different fixes, yet current benchmarks often merge them under task success, refusal, or final harmful outcome. We address this problem with PhoneSafety, a benchmark of 700 safety-critical moments drawn from real phone interactions across more than 130 apps. Each instance isolates the next decision at a risky moment and asks a simple question: does the model take the safe action, take the unsafe action, or fail to do anything useful? We evaluate eight representative phone-use agents under this framework. Our results reveal two main patterns. First, stronger general phone-use ability does not reliably imply safer choices at risky moments. Models that perform better on ordinary app tasks are not always the ones that behave more safely when the next action matters. Second, failures to do anything useful behave like a capability signal rather than a safety signal: they are concentrated in more visually and operationally demanding settings and remain stable when the evaluation protocol changes. Across models, failures split into two recurring patterns: unsafe choices in settings where the model can act but chooses wrongly, and inability to act in more visually and operationally demanding screens. Overall, a harmless outcome is not enough to count as evidence of safety. Evaluating phone-use agents requires separating unsafe judgment from inability to act.

The expansion of text-to-image diffusion models has raised concerns about harmful outputs, from fabricated depictions of public figures to sexually explicit imagery. To mitigate such risks, prior work has proposed concept erasure methods that aim to sever unwanted concepts from the model via fine-tuning, yet it remains unclear whether these approaches truly remove all links to the harmful concept or merely conceal superficial connections. In this work, we reveal a critical vulnerability, the Erasure Evasion Backdoor (EEB): an adversary binds a backdoor trigger to a concept slated for removal, and this malicious link survives subsequent erasure. We show that both black-box and white-box adversaries can instantiate this threat. Across six state-of-the-art erasure methods, including robust ones that explicitly search for alternative representations of the target concept, EEB consistently exposes harmful content: up to 82% success against celebrity-identity unlearning, up to 94% for object erasure, and up to 16 times amplification of explicit-content exposure. While EEB uncovers a blind spot in current erasure methods, it also provides a diagnostic tool for stress-testing future concept erasure techniques.

AI agents extend large language models (LLMs) with external tools, enabling them to perform complex tasks and translate model outputs into consequential real-world actions. Yet LLMs often become substantially less safe when deployed as agents, and the source of this degradation remains poorly understood. In this paper, we identify schema-formatted tool specifications as a primary source of agent safety degradation and show, through white-box representation analysis, that they weaken the model's internal refusal signals and contribute to unsafe tool execution. Building on this finding, we propose SafeKeep, an inference-time safeguard that decouples safety judgment from tool execution: it assesses requests using flattened textual tool specifications while retaining the original schema-formatted specifications for execution. Across two representative benchmarks and four LLMs, including both white-box and black-box models, SafeKeep increases the average refusal rate for harmful requests from 23.8% to 70.6% and reduces the average attack success rate under observation-level prompt injection from 25.6% to 2.5%. It also outperforms existing safeguards and preserves task-handling capability. We release the code and data at https://github.com/snowcatsmoking/SafeKeep .

World Action Models (WAMs) augment robot policies with action-conditioned predicted futures, but a plausible future alone does not justify changing the action that a bimanual policy would execute. We present CoWAM, a selective intervention layer that expresses synchronization, role compatibility, and collision convergence as coordination contracts. Each contract combines typed admissibility checks with event-conditioned verification and calibrated intervention gates. CoWAM preserves the nominal action unless an alternative satisfies every active obligation and provides a clear, low-risk improvement; when the nominal action is also inadmissible, it invokes a predefined abstention fallback. To separate selector quality from proposal quality, all methods operate on identical candidate pools and commit their decisions before shared oracle labeling. Across eight simulated bimanual tasks, CoWAM improves coordination-valid selection by 16.7 percentage points over the contract-only variant and raises closed-loop success by 9.6 percentage points over the strongest selective baseline, while keeping harmful interventions below 1%. Together, these results establish coordination contracts as an effective interface for conservative policy intervention with predicted world-action evidence across coordination-rich bimanual tasks.

While poisoning attacks on machine learning models have been extensively studied, the mechanisms by which adversaries can distribute poisoned models at scale remain largely unexplored. In this paper, we shed light on how model leaderboards -- ranked platforms for model discovery and evaluation -- can serve as a powerful channel for adversaries for stealthy large-scale distribution of poisoned models. We present TrojanClimb, a general framework that enables injection of malicious behaviors while maintaining competitive leaderboard performance. We demonstrate its effectiveness across four diverse modalities: text-embedding, text-generation, text-to-speech and text-to-image, showing that adversaries can successfully achieve high leaderboard rankings while embedding arbitrary harmful functionalities, from backdoors to bias injection. Our findings reveal a significant vulnerability in the machine learning ecosystem, highlighting the urgent need to redesign leaderboard evaluation mechanisms to detect and filter malicious (e.g., poisoned) models, while exposing broader security implications for the machine learning community regarding the risks of adopting models from unverified sources.

Recent reasoning-based safety guardrails for Large Reasoning Models (LRMs), such as deliberative alignment, have shown strong defense against jailbreak attacks. By leveraging LRMs' reasoning ability, these guardrails help the models to assess the safety of user inputs before generating final responses. The powerful reasoning ability can analyze the intention of the input query and will refuse to assist once it detects the harmful intent hidden by the jailbreak methods. Such guardrails have shown a significant boost in defense, such as the near-perfect refusal rates on the open-source gpt-oss series. Unfortunately, we find that these powerful reasoning-based guardrails can be extremely vulnerable to subtle manipulation of the input prompts, and once hijacked, can lead to even more harmful results. Specifically, we first uncover a surprisingly fragile aspect of these guardrails: simply adding a few template tokens to the input prompt can successfully bypass the seemingly powerful guardrails and lead to explicit and harmful responses. To explore further, we introduce a bag of jailbreak methods that subvert the reasoning-based guardrails. Our attacks span white-, gray-, and black-box settings and range from effortless template manipulations to fully automated optimization. Along with the potential for scalable implementation, these methods also achieve alarmingly high attack success rates (e.g., exceeding 90% across 5 different benchmarks on gpt-oss series on both local host models and online API services). Evaluations across various leading open-source LRMs confirm that these vulnerabilities are systemic, underscoring the urgent need for stronger alignment techniques for open-sourced LRMs to prevent malicious misuse. Code is open-sourced at https://chenxshuo.github.io/bag-of-tricks.

Model merging has emerged as an efficient and flexible paradigm for multi-task learning, with numerous methods being proposed in recent years. However, these state-of-the-art techniques are typically evaluated on benchmark suites that are highly favorable to model merging, and their robustness in more realistic settings remains largely unexplored. In this work, we first investigate the vulnerabilities of model-merging methods and pinpoint the source-model characteristics that critically underlie them. Specifically, we identify two factors that are particularly harmful to the merging process: (1) disparities in task vector norms, and (2) the low confidence of the source models. To address this issue, we propose DisTaC (Distillation for Task vector Conditioning), a novel method that pre-conditions these problematic task vectors before the merge. DisTaC leverages knowledge distillation to adjust a task vector's norm and increase source-model confidence while preserving its essential task-specific knowledge. Our extensive experiments demonstrate that by pre-conditioning task vectors with DisTaC, state-of-the-art merging techniques can successfully integrate models exhibiting the harmful traits -- where they would otherwise fail -- achieving significant performance gains.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a promising approach to improve the reasoning abilities of Large Language Models (LLMs). Among RLVR algorithms, Group Relative Policy Optimization (GRPO) and its variants have demonstrated strong performance and high training efficiency. However, GRPO-style objectives exhibit two issues on high accuracy prompts including mastered prompts (rollout accuracy =1) and majority-correct prompts (rollout accuracy in (0.5,1)). For mastered prompts, group-relative advantages vanish, yielding no training signal and unconstrained policy drift that can cause forgetting. For majority-correct prompts, the induced query weight shrinks as accuracy increases, weakening consolidation from partial correctness to mastery. To alleviate this, we propose Mastery-Consolidated Policy Optimization (MCPO), which introduces (i) a hinge-KL regularizer applied exclusively to mastered prompts to bound harmful policy drift between successive gradient steps, and (ii) a weighting mechanism that prioritizes majority-correct prompts to better allocate optimization effort. Extensive experiments across three mathematical benchmarks demonstrate that MCPO consistently improves pass@1 performance. Counter-intuitively, rather than restricting exploration, MCPO boosts pass@k metrics, indicating that mastery consolidation further catalyzes solution diversity.

Large Language Models (LLMs) face prominent security risks from jailbreaking, a practice that manipulates models to bypass built-in security constraints and generate unethical or unsafe content. Among various jailbreak techniques, multi-turn jailbreak attacks are more covert and persistent than single-turn counterparts, exposing critical vulnerabilities of LLMs. However, existing multi-turn jailbreak methods suffer from two fundamental limitations that affect the actual impact in real-world scenarios: (a) As models become more context-aware, any explicit harmful trigger is increasingly likely to be flagged and blocked; (b) Successful final-step triggers often require finely tuned, model-specific contexts, making such attacks highly context-dependent. To fill this gap, we propose Salami Slicing Risk, which operates by chaining numerous low-risk inputs that individually evade alignment thresholds but cumulatively accumulate harmful intent to ultimately trigger high-risk behaviors, without heavy reliance on pre-designed contextual structures. Building on this risk, we develop Salami Attack, an automatic framework universally applicable to multiple model types and modalities. Rigorous experiments demonstrate its state-of-the-art performance across diverse models and modalities, achieving over 90\% Attack Success Rate on GPT-4o and Gemini, as well as robustness against real-world alignment defenses. We also proposed a defense strategy to constrain the Salami Attack by at least 44.8\% while achieving a maximum blocking rate of 64.8\% against other multi-turn jailbreak attacks. Our findings provide critical insights into the pervasive risks of multi-turn jailbreaking and offer actionable mitigation strategies to enhance LLM security.

LLM developers have imposed technical interventions to prevent fine-tuning misuse attacks, attacks where adversaries evade safeguards by fine-tuning the model using a public API. Previous work has established several successful attacks against specific fine-tuning API defences. In this work, we show that defences of fine-tuning APIs that seek to detect individual harmful training or inference samples ('pointwise' detection) are fundamentally limited in their ability to prevent fine-tuning attacks. We construct 'pointwise-undetectable' attacks that repurpose entropy in benign model outputs (e.g. semantic or syntactic variations) to covertly transmit dangerous knowledge. Our attacks are composed solely of unsuspicious benign samples that can be collected from the model before fine-tuning, meaning training and inference samples are all individually benign and low-perplexity. We test our attacks against the OpenAI fine-tuning API, finding they succeed in eliciting answers to harmful multiple-choice questions, and that they evade an enhanced monitoring system we design that successfully detects other fine-tuning attacks. We encourage the community to develop defences that tackle the fundamental limitations we uncover in pointwise fine-tuning API defences.

Safety-aligned language models often refuse a harmful request stated directly but answer the same request inside a role-play or narrative wrapper. We measure this vulnerability across languages and registers: attack success on Qwen3-1.7B is already 89.4% in English and 93.0% in modern Chinese, and reaches 95.7% in Classical Chinese. We build GUISE, a benchmark for systematically studying this vulnerability. It includes parallel requests in English, modern Chinese, and Classical Chinese, matched harmful and benign pairs, wrapper types held out for evaluation, and a stricter criterion that counts warn-then-answer responses as attack successes. Representation analysis shows that language and register move harmful-request representations only slightly away from the model's refusal direction, whereas narrative wrappers move them much farther away. We propose AXIS, which combines preference optimisation with a rotation objective that aligns harmful-request representations with the refusal direction and a commitment objective that trains the model to refuse completely rather than produce a warn-then-answer response. Across Qwen3-1.7B, Qwen3-4B and GLM-4-9B, AXIS achieves the highest combined safety and usability score among the compared methods.

Safety aligned Large Language Models (LLMs) are vulnerable to harmful fine-tuning attacks qi2023fine-- a few harmful data mixed in the fine-tuning dataset can break the LLMs's safety alignment. Existing mitigation strategies include alignment stage solutions huang2024vaccine, rosati2024representation and fine-tuning stage solutions huang2024lazy,mukhoti2023fine. However, our evaluation shows that both categories of defenses fail when some specific training hyper-parameters are chosen -- a large learning rate or a large number of training epochs in the fine-tuning stage can easily invalidate the defense, which however, is necessary to guarantee finetune performance. To this end, we propose Antidote, a post-fine-tuning stage solution, which remains \textit{agnostic to the training hyper-parameters in the fine-tuning stage}. Antidote relies on the philosophy that by removing the harmful parameters, the harmful model can be recovered from the harmful behaviors, regardless of how those harmful parameters are formed in the fine-tuning stage. With this philosophy, we introduce a one-shot pruning stage after harmful fine-tuning to remove the harmful weights that are responsible for the generation of harmful content. Despite its embarrassing simplicity, empirical results show that Antidote can reduce harmful score while maintaining accuracy on downstream tasks.Our project page is at https://huangtiansheng.github.io/Antidote_gh_page/

Generative language models (LMs) offer numerous advantages but may produce inappropriate or harmful outputs due to the harmful knowledge acquired during pre-training. This knowledge often manifests as undesirable correspondences, such as "harmful prompts" leading to "harmful outputs," which our research aims to mitigate through unlearning techniques.However, existing unlearning methods based on gradient ascent can significantly impair the performance of LMs. To address this issue, we propose a novel approach called Weighted Positional N-pair (WPN) Learning, which leverages position-weighted mean pooling within an n-pair contrastive learning framework. WPN is designed to modify the output distribution of LMs by eliminating specific harmful outputs (e.g., replacing toxic responses with neutral ones), thereby transforming the model's behavior from "harmful prompt-harmful output" to "harmful prompt-harmless response".Experiments on OPT and GPT-NEO LMs show that WPN effectively reduces the proportion of harmful responses, achieving a harmless rate of up to 95.8\% while maintaining stable performance on nine common benchmarks (with less than 2\% degradation on average). Moreover, we provide empirical evidence to demonstrate WPN's ability to weaken the harmful correspondences in terms of generalizability and robustness, as evaluated on out-of-distribution test sets and under adversarial attacks.

Frontier large language models (LLMs) safety evaluation has largely treated harmful generation as an attack outcome rather than as an object of analysis. Consequently, little is known about the harmful outputs produced during model misbehavior, partly because large-scale, high-quality collections of frontier-LLM misbehavior are difficult to obtain. To address this gap, we introduce HarmProfile, a content-centric benchmark dataset that collects model misbehavior across diverse harm categories and model families, and defines the resulting harmful-output distribution as a model-level risk profile. The premise is that, just as linguistic behavior can be characterized from an utterance corpus, model risk can be characterized from the content, severity, and variation of its safety failures. HarmProfile contains over 80,000 validated artifacts from 23 frontier LLMs across 13 model families, organized into 15 harm categories and 57 subcategories. Using this corpus, we find that frontier LLMs reliably produce harmful content at scale, yet exhibit distinct risk profiles; both harmfulness and diversity grow with model capability, suggesting that frontier LLMs may appear safe yet harbor increasingly dangerous knowledge beneath the alignment surface. Our source code is available at https://github.com/fresh-ma/HarmProfile .

Current approaches to building general-purpose AI systems tend to produce systems with both beneficial and harmful capabilities. Further progress in AI development could lead to capabilities that pose extreme risks, such as offensive cyber capabilities or strong manipulation skills. We explain why model evaluation is critical for addressing extreme risks. Developers must be able to identify dangerous capabilities (through "dangerous capability evaluations") and the propensity of models to apply their capabilities for harm (through "alignment evaluations"). These evaluations will become critical for keeping policymakers and other stakeholders informed, and for making responsible decisions about model training, deployment, and security.

Large language models (LLMs) have evolved into autonomous agents that rely on open skill ecosystems (e.g., ClawHub and Skills.Rest), hosting numerous publicly reusable skills. Existing security research on these ecosystems mainly focuses on vulnerabilities within skills, such as prompt injection. However, there is a critical gap regarding skills that may be misused for harmful actions (e.g., cyber attacks, fraud and scams, privacy violations, and sexual content generation), namely harmful skills. In this paper, we present the first large-scale measurement study of harmful skills in agent ecosystems, covering 98,440 skills across two major registries. Using an LLM-driven scoring system grounded in our harmful skill taxonomy, we find that 4.93% of skills (4,858) are harmful, with ClawHub exhibiting an 8.84% harmful rate compared to 3.49% on Skills.Rest. We then construct HarmfulSkillBench, the first benchmark for evaluating agent safety against harmful skills in realistic agent contexts, comprising 200 harmful skills across 20 categories and four evaluation conditions. By evaluating six LLMs on HarmfulSkillBench, we find that presenting a harmful task through a pre-installed skill substantially lowers refusal rates across all models, with the average harm score rising from 0.27 without the skill to 0.47 with it, and further to 0.76 when the harmful intent is implicit rather than stated as an explicit user request. We responsibly disclose our findings to the affected registries and release our benchmark to support future research (see https://github.com/TrustAIRLab/HarmfulSkillBench).

We study how to perform unlearning, i.e. forgetting undesirable misbehaviors, on large language models (LLMs). We show at least three scenarios of aligning LLMs with human preferences can benefit from unlearning: (1) removing harmful responses, (2) erasing copyright-protected content as requested, and (3) reducing hallucinations. Unlearning, as an alignment technique, has three advantages. (1) It only requires negative (e.g. harmful) examples, which are much easier and cheaper to collect (e.g. via red teaming or user reporting) than positive (e.g. helpful and often human-written) examples required in RLHF (RL from human feedback). (2) It is computationally efficient. (3) It is especially effective when we know which training samples cause the misbehavior. To the best of our knowledge, our work is among the first to explore LLM unlearning. We are also among the first to formulate the settings, goals, and evaluations in LLM unlearning. We show that if practitioners only have limited resources, and therefore the priority is to stop generating undesirable outputs rather than to try to generate desirable outputs, unlearning is particularly appealing. Despite only having negative samples, our ablation study shows that unlearning can still achieve better alignment performance than RLHF with just 2% of its computational time.

Imitation-learning policies for robot manipulation inherit the quality of the success labels attached to their training episodes, and those labels are usually produced by the robot's own success check. A particularly damaging error is the false success: an episode the robot logs as a success when the task outcome was actually wrong. We ask a narrow but practical question about these episodes. Once an episode has already been flagged as a success, how much of the information needed to overturn that label is present in proprioception, and how much requires vision? We build a simulated testbed on two bimanual ALOHA tasks, induce failures through environment perturbations rather than label edits, label every episode by privileged simulator state that the detector never sees, and keep only episodes the robot flagged as successful. We then compare detectors restricted to proprioception against a vision-based detector. We find that recoverability spans a wide range: in cube transfer the false successes are almost fully recoverable from joint data alone, while in peg insertion proprioception recovers only part of them and a vision detector closes most of the gap. We also show that the proprioceptive separability we measure rests on velocity differences far below any realistic sensor noise floor, so it is best read as an optimistic upper bound that a noiseless simulator inflates. We release the generation and evaluation pipeline.

Language models are increasingly capable and are being rapidly deployed on a population-level scale. As a result, the safety of these models is increasingly high-stakes. Fortunately, advances in alignment have significantly reduced the likelihood of harmful model outputs. However, when models are queried billions of times in a day, even rare worst-case behaviors will occur. Current safety evaluations focus on capturing the distribution of inputs that yield harmful outputs. These evaluations disregard the probabilistic nature of models and their tail output behavior. To measure this tail risk, we propose a method to efficiently estimate the probability of harmful outputs for any input query. Instead of naive brute-force sampling from the target model, where harmful outputs could be rare, we operationalize importance sampling by creating unsafe versions of the target model. These unsafe versions enable sample-efficient estimation by making harmful outputs more probable. On benchmarks measuring misuse and misalignment, these estimates match brute-force Monte Carlo estimates using 10-20x fewer samples. For example, we can estimate probability of harmful outputs on the order of 10^-4 with just 500 samples. Additionally, we find that these harmfulness estimates can reveal the sensitivity of models to perturbations in model input and predict deployment risks. Our work demonstrates that accurate rare-event estimation is both critical and feasible for safety evaluations. Code is available at https://github.com/rangell/LMTailRisk

Recent advances in large language models (LLMs) have intensified the need to understand and reliably curb their harmful behaviours. We introduce a multidimensional framework for probing and steering harmful content in model internals. For each of 55 distinct harmfulness subconcepts (e.g., racial hate, employment scams, weapons), we learn a linear probe, yielding 55 interpretable directions in activation space. Collectively, these directions span a harmfulness subspace that we show is strikingly low-rank. We then test ablation of the entire subspace from model internals, as well as steering and ablation in the subspace's dominant direction. We find that dominant direction steering allows for near elimination of harmfulness with a low decrease in utility. Our findings advance the emerging view that concept subspaces provide a scalable lens on LLM behaviour and offer practical tools for the community to audit and harden future generations of language models.