Daily Papers

Large language models (LLMs) can serve as helpful assistants, yet they can equally function as harm amplifiers that enable malicious users to achieve harmful outcomes beyond their capabilities through extended interactions. This risk manifests along two axes, i.e., democratizing domain expertise that allows novices to produce specialized harmful content, and scaling harmful operations at volumes that manual effort cannot match. Existing works, however, often overlook how LLMs compound harm across multi-turn conversations. We introduce HarmAmp, a new benchmark for multi-turn harm amplification scenarios spanning twelve risk categories. Each scenario is grounded in real-world threats and satisfies rigorous criteria, i.e., substantive amplification, operational specificity, and multi-turn necessity. We further propose TrajSafe, a proactive monitor that anticipates harmful trajectories and intervenes through actions such as probing users' genuine intents and steering the models towards safer completion. Our extensive experiments demonstrate that TrajSafe significantly reduces the harmfulness incurred in multi-turn interactions while preserving a low over-refusal rate and the target model's general capabilities. Our work offers a promising paradigm to alleviate the nuanced safety risks in LLM interactions.

Large Language Models (LLMs) are improving at an exceptional rate. With the advent of agentic workflows, multi-turn dialogue has become the de facto mode of interaction with LLMs for completing long and complex tasks. While LLM capabilities continue to improve, they remain increasingly susceptible to jailbreaking, especially in multi-turn scenarios where harmful intent can be subtly injected across the conversation to produce nefarious outcomes. While single-turn attacks have been extensively explored, adaptability, efficiency and effectiveness continue to remain key challenges for their multi-turn counterparts. To address these gaps, we present PLAGUE, a novel plug-and-play framework for designing multi-turn attacks inspired by lifelong-learning agents. PLAGUE dissects the lifetime of a multi-turn attack into three carefully designed phases (Primer, Planner and Finisher) that enable a systematic and information-rich exploration of the multi-turn attack family. Evaluations show that red-teaming agents designed using PLAGUE achieve state-of-the-art jailbreaking results, improving attack success rates (ASR) by more than 30% across leading models in a lesser or comparable query budget. Particularly, PLAGUE enables an ASR (based on StrongReject) of 81.4% on OpenAI's o3 and 67.3% on Claude's Opus 4.1, two models that are considered highly resistant to jailbreaks in safety literature. Our work offers tools and insights to understand the importance of plan initialization, context optimization and lifelong learning in crafting multi-turn attacks for a comprehensive model vulnerability evaluation.

LLM-based agents are becoming increasingly capable, yet their safety lags behind. This creates a gap between what agents can do and should do. This gap widens as agents engage in multi-turn interactions and employ diverse tools, introducing new risks overlooked by existing benchmarks. To systematically scale safety testing into multi-turn, tool-realistic settings, we propose a principled taxonomy that transforms single-turn harmful tasks into multi-turn attack sequences. Using this taxonomy, we construct MT-AgentRisk (Multi-Turn Agent Risk Benchmark), the first benchmark to evaluate multi-turn tool-using agent safety. Our experiments reveal substantial safety degradation: the Attack Success Rate (ASR) increases by 16% on average across open and closed models in multi-turn settings. To close this gap, we propose ToolShield, a training-free, tool-agnostic, self-exploration defense: when encountering a new tool, the agent autonomously generates test cases, executes them to observe downstream effects, and distills safety experiences for deployment. Experiments show that ToolShield effectively reduces ASR by 30% on average in multi-turn interactions. Our code is available at https://github.com/CHATS-lab/ToolShield.

Hidden malicious intent in multi-turn dialogue poses a growing threat to deployed large language models (LLMs). Rather than exposing a harmful objective in a single prompt, increasingly capable attackers can distribute their intent across multiple benign-looking turns. Recent studies show that even modern commercial models with advanced guardrails remain vulnerable to such attacks despite advances in safety alignment and external guardrails. In this work, we address this challenge by detecting the earliest turn at which delivering the candidate response would make the accumulated interaction sufficient to enable harmful action. This objective requires precise turn-level intervention that identifies the harm-enabling closure point while avoiding premature refusal of benign exploratory conversations. To further support training and evaluation, we construct the Multi-Turn Intent Dataset (MTID), which contains branching attack rollouts, matched benign hard negatives, and annotations of the earliest harm-enabling turns. We show that MTID helps enable a turn-level monitor TurnGate, which substantially outperforms existing baselines in harmful-intent detection while maintaining low over-refusal rates. TurnGate further generalizes across domains, attacker pipelines, and target models. Our code is available at https://github.com/Graph-COM/TurnGate.

Despite extensive safety enhancements in large language models (LLMs), multi-turn "jailbreak" conversations crafted by skilled human adversaries can still breach even the most sophisticated guardrails. However, these multi-turn attacks demand considerable manual effort, limiting their scalability. In this work, we introduce a novel approach called Multi-turn-to-Single-turn (M2S) that systematically converts multi-turn jailbreak prompts into single-turn attacks. Specifically, we propose three conversion strategies - Hyphenize, Numberize, and Pythonize - each preserving sequential context yet packaging it in a single query. Our experiments on the Multi-turn Human Jailbreak (MHJ) dataset show that M2S often increases or maintains high Attack Success Rates (ASRs) compared to original multi-turn conversations. Notably, using a StrongREJECT-based evaluation of harmfulness, M2S achieves up to 95.9% ASR on Mistral-7B and outperforms original multi-turn prompts by as much as 17.5% in absolute improvement on GPT-4o. Further analysis reveals that certain adversarial tactics, when consolidated into a single prompt, exploit structural formatting cues to evade standard policy checks. These findings underscore that single-turn attacks - despite being simpler and cheaper to conduct - can be just as potent, if not more, than their multi-turn counterparts. Our findings underscore the urgent need to reevaluate and reinforce LLM safety strategies, given how adversarial queries can be compacted into a single prompt while still retaining sufficient complexity to bypass existing safety measures.

We present MultiBreak, a scalable and diverse multi-turn jailbreak benchmark to evaluate large language model (LLM) safety. Multi-turn jailbreaks mimic natural conversational settings, making them easier to bypass safety-aligned LLM than single-turn jailbreaks. Existing multi-turn benchmarks are limited in size or rely heavily on templates, which restrict their diversity. To address this gap, we unify a wide range of harmful jailbreak intents, and introduce an active learning pipeline for expanding high-quality multi-turn adversarial prompts, where a generator is iteratively fine-tuned to produce stronger attack candidates, guided by uncertainty-based refinement. Our MultiBreak includes 10,389 multi-turn adversarial prompts, spans 2,665 distinct harmful intents, and covers the most diverse set of topics to date. Empirical evaluation shows that our benchmark achieves up to a 54.0 and 34.6 higher attack success rate (ASR)} than the second-best dataset on DeepSeek-R1-7B and GPT-4.1-mini, respectively. More importantly, safety evaluations suggest that diverse attack categories uncover fine-grained LLM vulnerabilities}, and categories that appear benign under single-turn can exhibit substantially higher adversarial effectiveness in multi-turn scenarios. These findings highlight persistent vulnerabilities of LLMs under realistic adversarial settings and establish MultiBreak as a scalable resource for advancing LLM safety.

This study exposes the safety vulnerabilities of Large Language Models (LLMs) in multi-turn interactions, where malicious users can obscure harmful intents across several queries. We introduce ActorAttack, a novel multi-turn attack method inspired by actor-network theory, which models a network of semantically linked actors as attack clues to generate diverse and effective attack paths toward harmful targets. ActorAttack addresses two main challenges in multi-turn attacks: (1) concealing harmful intents by creating an innocuous conversation topic about the actor, and (2) uncovering diverse attack paths towards the same harmful target by leveraging LLMs' knowledge to specify the correlated actors as various attack clues. In this way, ActorAttack outperforms existing single-turn and multi-turn attack methods across advanced aligned LLMs, even for GPT-o1. We will publish a dataset called SafeMTData, which includes multi-turn adversarial prompts and safety alignment data, generated by ActorAttack. We demonstrate that models safety-tuned using our safety dataset are more robust to multi-turn attacks. Code is available at https://github.com/renqibing/ActorAttack.

Deploying LLMs in multi-turn dialogues facilitates jailbreak attacks that distribute harmful intent across seemingly benign turns. Recent training-based multi-turn jailbreak methods learn long-horizon attack strategies from interaction feedback, but often rely on coarse trajectory-level outcome signals that broadcast uniformly to every turn. However, we find that turn-level contributions in multi-turn jailbreaking are non-uniform, phase-dependent, and target-specific. Such coarse outcome supervision induces a credit assignment problem, leading to over-rewarding redundant turns in successful trajectories and under-crediting useful intermediate turns in failed ones. To address this, we propose TRACE, a turn-aware credit assignment framework for reinforcement learning (RL)-based multi-turn jailbreaking. For successful trajectories, TRACE estimates turn-level contributions via leave-one-turn-out semantic masking; for failed ones, TRACE assigns penalties based on prompt harmfulness and semantic relevance, with an additional local refusal-aware penalty. Furthermore, we reuse the attack-side credit signal for multi-turn defense alignment. Extensive experiments on open-source and closed-source targets show that TRACE achieves strong overall performance in effectiveness, transferability, and efficiency, yielding about a 25% relative improvement in attack success rate over the strongest RL baseline while also improving the safety-utility balance when reused for defense alignment.

Multi-turn jailbreak attacks have emerged as a critical safety threat to LLMs, as harmful objectives are decomposed across a sequence of apparently benign turns to bypass guardrails. Existing defenses lack the reasoning capacity to identify evolving manipulation patterns, often trading helpfulness for safety by over-refusing benign requests related to sensitive topics. We introduce Trace, a multi-turn defense with trajectory-aware structured reasoning. Before generating each response, the model identifies manipulation cues from the trajectory, evaluates both the benign and adversarial interpretations of user intent, assigns a jailbreak score, and commits to an action: Allow, Caution, or Decline. We curate 4k multi-turn adversarial conversations from five attack frameworks, pair them with 2.4k benign dialogs, and 600 sensitive-but-benign conversations. We train Llama-3.1-8B-Instruct with SFT and GRPO under a multi-component reward that jointly optimizes helpfulness on benign prompts and robustness against jailbreak attempts. Across seven multi-turn attack benchmarks, Trace attains an average attack success rate (ASR) of 14.5% against 31.4% for the strongest baseline and 74.9% for the undefended target, while significantly raising the attacker effort required per successful jailbreak. Trace also balances usability and safety, achieving a 93.3% average compliance on over-refusal benchmarks.

Recent large language model (LLM) defenses have greatly improved models' ability to refuse harmful queries, even when adversarially attacked. However, LLM defenses are primarily evaluated against automated adversarial attacks in a single turn of conversation, an insufficient threat model for real-world malicious use. We demonstrate that multi-turn human jailbreaks uncover significant vulnerabilities, exceeding 70% attack success rate (ASR) on HarmBench against defenses that report single-digit ASRs with automated single-turn attacks. Human jailbreaks also reveal vulnerabilities in machine unlearning defenses, successfully recovering dual-use biosecurity knowledge from unlearned models. We compile these results into Multi-Turn Human Jailbreaks (MHJ), a dataset of 2,912 prompts across 537 multi-turn jailbreaks. We publicly release MHJ alongside a compendium of jailbreak tactics developed across dozens of commercial red teaming engagements, supporting research towards stronger LLM defenses.

Safety evaluation and red-teaming of large language models remain predominantly text-centric, and existing frameworks lack the infrastructure to systematically test whether alignment generalizes to audio, image, and video inputs. We present MUSE (Multimodal Unified Safety Evaluation), an open-source, run-centric platform that integrates automatic cross-modal payload generation, three multi-turn attack algorithms (Crescendo, PAIR, Violent Durian), provider-agnostic model routing, and an LLM judge with a five-level safety taxonomy into a single browser-based system. A dual-metric framework distinguishes hard Attack Success Rate (Compliance only) from soft ASR (including Partial Compliance), capturing partial information leakage that binary metrics miss. To probe whether alignment generalizes across modality boundaries, we introduce Inter-Turn Modality Switching (ITMS), which augments multi-turn attacks with per-turn modality rotation. Experiments across six multimodal LLMs from four providers show that multi-turn strategies can achieve up to 90-100% ASR against models with near-perfect single-turn refusal. ITMS does not uniformly raise final ASR on already-saturated baselines, but accelerates convergence by destabilizing early-turn defenses, and ablation reveals that the direction of modality effects is model-family-specific rather than universal, underscoring the need for provider-aware cross-modal safety testing.

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.

Vision Large Language Models (VLLMs) represent a significant advancement in artificial intelligence by integrating image-processing capabilities with textual understanding, thereby enhancing user interactions and expanding application domains. However, their increased complexity introduces novel safety and ethical challenges, particularly in multi-modal and multi-turn conversations. Traditional safety evaluation frameworks, designed for text-based, single-turn interactions, are inadequate for addressing these complexities. To bridge this gap, we introduce the REVEAL (Responsible Evaluation of Vision-Enabled AI LLMs) Framework, a scalable and automated pipeline for evaluating image-input harms in VLLMs. REVEAL includes automated image mining, synthetic adversarial data generation, multi-turn conversational expansion using crescendo attack strategies, and comprehensive harm assessment through evaluators like GPT-4o. We extensively evaluated five state-of-the-art VLLMs, GPT-4o, Llama-3.2, Qwen2-VL, Phi3.5V, and Pixtral, across three important harm categories: sexual harm, violence, and misinformation. Our findings reveal that multi-turn interactions result in significantly higher defect rates compared to single-turn evaluations, highlighting deeper vulnerabilities in VLLMs. Notably, GPT-4o demonstrated the most balanced performance as measured by our Safety-Usability Index (SUI) followed closely by Pixtral. Additionally, misinformation emerged as a critical area requiring enhanced contextual defenses. Llama-3.2 exhibited the highest MT defect rate (16.55 %) while Qwen2-VL showed the highest MT refusal rate (19.1 %).

Despite substantial investment in safety alignment, the vulnerability of large language models to sophisticated multi-turn adversarial attacks remains poorly characterized, and whether model scale or inference mode affects robustness is unknown. This study employed the TEMPEST multi-turn attack framework to evaluate ten frontier models from eight vendors across 1,000 harmful behaviors, generating over 97,000 API queries across adversarial conversations with automated evaluation by independent safety classifiers. Results demonstrated a spectrum of vulnerability: six models achieved 96% to 100% attack success rate (ASR), while four showed meaningful resistance, with ASR ranging from 42% to 78%; enabling extended reasoning on identical architecture reduced ASR from 97% to 42%. These findings indicate that safety alignment quality varies substantially across vendors, that model scale does not predict adversarial robustness, and that thinking mode provides a deployable safety enhancement. Collectively, this work establishes that current alignment techniques remain fundamentally vulnerable to adaptive multi-turn attacks regardless of model scale, while identifying deliberative inference as a promising defense direction.

As Vision-Language Models (VLMs) move into interactive, multi-turn use, new safety risks arise that single-turn or single-modality moderation misses. In Multimodal Multi-Turn (MMT) dialogues, malicious intent can be spread across turns and images, while context-sensitive replies may still advance harmful content. To address this challenge, we present the first systematic definition and study of MMT dialogue safety. Building on this formulation, we introduce the Multimodal Multi-turn Dialogue Safety (MMDS) dataset. We further develop an automated multimodal multi-turn red-teaming framework based on Monte Carlo Tree Search (MCTS) to generate unsafe multimodal multi-turn dialogues for MMDS. MMDS contains 4,484 annotated multimodal dialogue samples with fine-grained safety ratings, policy dimension labels, and evidence-based rationales for both users and assistants. Leveraging MMDS, we present LLaVAShield, a powerful tool that jointly detects and assesses risk in user inputs and assistant responses. Across comprehensive experiments, LLaVAShield consistently outperforms strong baselines on MMT content moderation tasks and under dynamic policy configurations, establishing new state-of-the-art results. We will publicly release the dataset and model to support future research.

Multi-turn interactions with language models (LMs) pose critical safety risks, as harmful intent can be strategically spread across exchanges. Yet, the vast majority of prior work has focused on single-turn safety, while adaptability and diversity remain among the key challenges of multi-turn red-teaming. To address these challenges, we present X-Teaming, a scalable framework that systematically explores how seemingly harmless interactions escalate into harmful outcomes and generates corresponding attack scenarios. X-Teaming employs collaborative agents for planning, attack optimization, and verification, achieving state-of-the-art multi-turn jailbreak effectiveness and diversity with success rates up to 98.1% across representative leading open-weight and closed-source models. In particular, X-Teaming achieves a 96.2% attack success rate against the latest Claude 3.7 Sonnet model, which has been considered nearly immune to single-turn attacks. Building on X-Teaming, we introduce XGuard-Train, an open-source multi-turn safety training dataset that is 20x larger than the previous best resource, comprising 30K interactive jailbreaks, designed to enable robust multi-turn safety alignment for LMs. Our work offers essential tools and insights for mitigating sophisticated conversational attacks, advancing the multi-turn safety of LMs.

Multi-turn jailbreaks capture the real threat model for safety-aligned chatbots, where single-turn attacks are merely a special case. Yet existing approaches break under exploration complexity and intent drift. We propose SEMA, a simple yet effective framework that trains a multi-turn attacker without relying on any existing strategies or external data. SEMA comprises two stages. Prefilling self-tuning enables usable rollouts by fine-tuning on non-refusal, well-structured, multi-turn adversarial prompts that are self-generated with a minimal prefix, thereby stabilizing subsequent learning. Reinforcement learning with intent-drift-aware reward trains the attacker to elicit valid multi-turn adversarial prompts while maintaining the same harmful objective. We anchor harmful intent in multi-turn jailbreaks via an intent-drift-aware reward that combines intent alignment, compliance risk, and level of detail. Our open-loop attack regime avoids dependence on victim feedback, unifies single- and multi-turn settings, and reduces exploration complexity. Across multiple datasets, victim models, and jailbreak judges, our method achieves state-of-the-art (SOTA) attack success rates (ASR), outperforming all single-turn baselines, manually scripted and template-driven multi-turn baselines, as well as our SFT (Supervised Fine-Tuning) and DPO (Direct Preference Optimization) variants. For instance, SEMA performs an average 80.1% ASR@1 across three closed-source and open-source victim models on AdvBench, 33.9% over SOTA. The approach is compact, reproducible, and transfers across targets, providing a stronger and more realistic stress test for large language model (LLM) safety and enabling automatic redteaming to expose and localize failure modes. Our code is available at: https://github.com/fmmarkmq/SEMA.

The rapid progress of Large Language Models (LLMs) has opened up new opportunities across various domains and applications; yet it also presents challenges related to potential misuse. To mitigate such risks, red teaming has been employed as a proactive security measure to probe language models for harmful outputs via jailbreak attacks. However, current jailbreak attack approaches are single-turn with explicit malicious queries that do not fully capture the complexity of real-world interactions. In reality, users can engage in multi-turn interactions with LLM-based chat assistants, allowing them to conceal their true intentions in a more covert manner. To bridge this gap, we, first, propose a new jailbreak approach, RED QUEEN ATTACK. This method constructs a multi-turn scenario, concealing the malicious intent under the guise of preventing harm. We craft 40 scenarios that vary in turns and select 14 harmful categories to generate 56k multi-turn attack data points. We conduct comprehensive experiments on the RED QUEEN ATTACK with four representative LLM families of different sizes. Our experiments reveal that all LLMs are vulnerable to RED QUEEN ATTACK, reaching 87.62% attack success rate on GPT-4o and 75.4% on Llama3-70B. Further analysis reveals that larger models are more susceptible to the RED QUEEN ATTACK, with multi-turn structures and concealment strategies contributing to its success. To prioritize safety, we introduce a straightforward mitigation strategy called RED QUEEN GUARD, which aligns LLMs to effectively counter adversarial attacks. This approach reduces the attack success rate to below 1% while maintaining the model's performance across standard benchmarks. Full implementation and dataset are publicly accessible at https://github.com/kriti-hippo/red_queen.

Mitigating bias in machine learning systems requires refining our understanding of bias propagation pathways: from societal structures to large-scale data to trained models to impact on society. In this work, we focus on one aspect of the problem, namely bias amplification: the tendency of models to amplify the biases present in the data they are trained on. A metric for measuring bias amplification was introduced in the seminal work by Zhao et al. (2017); however, as we demonstrate, this metric suffers from a number of shortcomings including conflating different types of bias amplification and failing to account for varying base rates of protected attributes. We introduce and analyze a new, decoupled metric for measuring bias amplification, BiasAmp_{rightarrow} (Directional Bias Amplification). We thoroughly analyze and discuss both the technical assumptions and normative implications of this metric. We provide suggestions about its measurement by cautioning against predicting sensitive attributes, encouraging the use of confidence intervals due to fluctuations in the fairness of models across runs, and discussing the limitations of what this metric captures. Throughout this paper, we work to provide an interrogative look at the technical measurement of bias amplification, guided by our normative ideas of what we want it to encompass. Code is located at https://github.com/princetonvisualai/directional-bias-amp

Existing LLM safety evaluations rely on binary attack-success rates and domain-agnostic taxonomies, leaving regulated Banking, Financial Services, and Insurance (BFSI) deployments exposed to failures elicited through legally or professionally plausible framing. We introduce RAHS (Risk-Adjusted Harm Score), a risk-sensitive metric jointly capturing disclosure severity, disclaimer mitigation, and inter-judge agreement, and FinRedTeamBench, a 989-prompt benchmark spanning seven BFSI risk areas and 34 sub-categories mapped to regulatory frameworks. Evaluation uses an ensemble of three heterogeneous LLM judges, validated against human experts, and an adaptive multi-turn red-teaming pipeline. On nine open-weight models, RAHS preserves separation under near-ceiling ASR, ranking is stable under hyperparameter sweeps, and multi-turn pressure drives not only more jailbreaks but more operationally severe disclosures, exposing failure modes that single-turn, domain-agnostic evaluations cannot reveal.

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.

Large language models have seen widespread adoption, yet they remain vulnerable to multi-turn jailbreak attacks, threatening their safe deployment. This has led to the task of training automated multi-turn attackers to probe model safety vulnerabilities. However, existing approaches typically rely on turn-level optimization, which is insufficient for learning long-term attack strategies. To bridge this gap, we formulate this task as a multi-turn reinforcement learning problem, directly optimizing the harmfulness of the final-turn response as the outcome reward. To address the sparse supervision of the outcome reward, we introduce TROJail, which employs two process rewards to evaluate the utility of intermediate prompts and integrate them into advantage estimation. These rewards (1) penalize overly harmful prompts that trigger the model's refusal mechanism, and (2) encourage steering the semantic relevance of responses toward the targeted harmful content. Experimental results show improved attack success rates across multiple models and benchmarks, highlighting the effectiveness of our approach. The code is available at https://github.com/xxiqiao/TROJail. Warning: This paper contains examples of harmful content.

Multi-turn jailbreak attacks pose a growing threat to large language model (LLM) safety because they exploit feedback from auxiliary judge models to iteratively refine prompts toward harmful goals. Existing defenses largely detect or block unsafe content at individual turns or at the final response, leaving the judge-driven refinement loop intact and allowing attackers to extract informative feedback from intermediate interactions. We introduce D-Judge, a semantics-preserving output rewriting defense that intervenes directly in this loop by rewriting the victim LLM's responses before they are evaluated by the attacker's judge. By misaligning the judge's feedback signal without changing the meaning of the original response, D-Judge derails the attacker's prompt-refinement process, causing subsequent queries to be optimized against a distorted signal of attack progress. To improve D-Judge's ability to produce such rewrites, we construct a dataset of semantically equivalent response pairs that induce different judge-assigned harmfulness scores, and use it for supervised fine-tuning followed by direct preference optimization. Experiments on HarmBench show that D-Judge reduces the success rate of state-of-the-art multi-turn jailbreaks while preserving performance on benign benchmarks.

MLLMs are increasingly deployed in multi-turn settings, where attackers can escalate unsafe intent through the evolving visual-text history and exploit long-context safety decay. Yet safety alignment is still dominated by single-turn data and fixed-template dialogues, leaving a mismatch between training and deployment. To bridge this gap, we propose SaFeR-Steer, a progressive multi-turn alignment framework that combines staged synthetic bootstrapping with tutor-in-the-loop GRPO to train a single student under adaptive, on-policy attacks. We also introduce Trajectory-Consistent Summative Reward (TCSR), which aggregates the historical minimum and average of turn rewards so that any low-quality turn affects the trajectory-level return. I. Dataset. We release STEER, a multi-turn multimodal safety dataset with STEER-SFT (12,934), STEER-RL (2,000), and STEER-Bench (3,227) dialogues spanning 2-10 turns. II. Experiment. Starting from Qwen2.5-VL-3B/7B, SaFeR-Steer substantially improves Safety/Helpfulness on both single-turn (48.30/45.86 rightarrow 81.84/70.77 for 3B; 56.21/60.32 rightarrow 87.89/77.40 for 7B) and multi-turn benchmarks (12.55/27.13 rightarrow 55.58/70.27 for 3B; 24.66/46.48 rightarrow 64.89/72.35 for 7B), shifting failures to later turns and yielding robustness beyond scaling alone. Code is available at https://anonymous.4open.science/r/SaFeR-Steer

Clinical reasoning demands multi-step interactions -- gathering patient history, ordering tests, interpreting results, and making safe treatment decisions -- yet a unified training environment provides the breadth of clinical domains and specialized tools to train generalizable medical AI agents through reinforcement learning remains elusive. We present a comprehensive empirical study of multi-turn agentic RL for medical AI, built on , a gymnasium-compatible environment spanning 10 clinical domains with 3.6K+ tasks, 135 domain-specific tools, and a knowledge base of 828K medical passages. Our analysis reveals that agentic multi-turn structure degrades into verbose single-turn monologues, characterized by monotonic length explosion and a simultaneous erosion of tool-use frequency. We characterize how this collapse, alongside distillation instability, stems from the misalignment of sparse terminal rewards with sequential clinical trajectories. We find that vanilla GRPO achieves strong final accuracy on some benchmarks but suffers from training instability, evidenced by significant oscillations in response length and prolonged convergence periods. To improve training efficiency and stability, we propose Turn-level Truncated On-Policy Distillation (TT-OPD), a self-distillation framework where a gradient-free EMA teacher leverages outcome-privileged information to provide dense, outcome-aware KL regularization at every conversation turn. TT-OPD achieves the best performance on 10 of 18 benchmarks with an average +3.9~pp improvement over the non-RL baseline with faster early convergence, controlled response length, and sustained multi-turn tool use.

Large language models (LLMs) are increasingly explored as scalable tools for mental health counseling, yet evaluating their safety remains challenging due to the interactional and context-dependent nature of clinical harm. Existing evaluation frameworks predominantly assess isolated responses using coarse-grained taxonomies or static datasets, limiting their ability to diagnose how harms emerge and accumulate over multi-turn counseling interactions. In this work, we introduce R-MHSafe, a role-aware mental health safety taxonomy that characterizes clinically significant harm in terms of the interactional roles an AI counselor adopts, including perpetrator, instigator, facilitator, or enabler, combined with clinically grounded harm categories. Then, we propose MHSafeEval, a closed-loop, agent-based evaluation framework that formulates safety assessment as trajectory-level discovery of harm through adversarial multi-turn interactions, guided by role-aware modeling. Using R-MHSafe and MHSafeEval, we conduct a large-scale evaluation across state-of-the-art LLMs. Our results reveal substantial role-dependent and cumulative safety failures that are systematically missed by existing static benchmarks, and show that our framework significantly improves failure-mode coverage and diagnostic granularity.

Language models have been adopted faster than evaluation practices that reliably assess their risks, and a growing share of use is affective rather than informational. Within that traffic users disclose suicide and self-harm (SSH) crises. Existing safety benchmarks score models primarily on refusal, but refusal is not the desired behavior here; the conversation requires a substantive supportive response. We introduce DistressBench, 718 English conversations with clinician-authored user turns spanning 23 SSH subcategories, each paired with a task-specific rubric of weighted atomic criteria across seven dimensions of crisis support and a clinician reference response. A model's score is the weighted average over the criteria its reply satisfies, reading directly as the fraction of required care delivered. The best of 25 frontier models satisfies 88.4% of weighted criteria and the median model 71.4%, against 98.7% for clinician references scored on the same criteria. Pooled across all 25 models, in 35.3% of conversations where a model satisfied every criterion for recognizing the crisis, it failed to route the user toward human help. The shortfall concentrates in the dimensions requiring the model to act rather than withhold. Contrasting each dimension's median model against its best separates capability limits from choices: de-escalation reaches 74.6% in the strongest model against a median of 46.0%, and safety disclaimers 84.6% against 8.3%. Multi-turn conversations score below single-turn ones (68.5% against 74.2%). We release the corpus, rubrics, judge calibration labels, and a companion annex of 20 long-horizon red-team transcripts. These results argue for evaluating crisis safety on the care a model delivers rather than on what it declines to say, and for treating de-escalation and resource routing as deployment requirements.

Egocentric multimodal assistants see through the user's eyes and guide what the user does next, so their outputs directly shape physical actions, including those that affect bystanders. Yet existing safety evaluations do not distinguish whom a response would harm, and score only whether a model refuses. Pooling all harms into one refusal rate hides two failures: a model may protect one target but not another, and it may refuse without understanding the risk, or understand it and comply anyway. We present STOU, a benchmark of 2,902 egocentric video-audio and video-text queries whose potential harm targets the user (Self), co-present individuals (The Others), or society at large (Others). Our evaluation decouples safety understanding from refusal behavior: beyond refusal, we score whether the model recognizes the user's intent, the presence and target of harm, and its downstream consequences. Across 10 open and closed source omnimodal and multimodal LLMs, we find that refusal varies systematically with the harm target, and that models often recognize a harm yet still assist with it. These results show that refusal rates alone cannot certify an assistant as safe, and that safety alignment must account for who would be harmed. Project page: https://onedayoneyeah.github.io/safety-for-whom

Failures in multi-turn reasoning models are largely invisible to terminal-score evaluation. A model can lock onto an unsafe stance early in a long dialogue, yet its final-turn refusal rate may appear indistinguishable from a robustly aligned baseline. To expose these hidden temporal dynamics, we propose a trace-level diagnostic - the CoT-Output 2x2 safety matrix. This framework labels every turn along two independent axes (internal reasoning and visible output), yielding four operationally defined failure cells: robust alignment, alignment faking, overt jailbreak, and a distinct failure mode we term context-injection failure (where the CoT maintains safe reasoning, but the visible output produces harm, highlighting a multi-turn manifestation of reasoning unfaithfulness). We evaluate three distilled reasoning targets against a fixed attacker across five oversight conditions, collecting 6750 turn-level observations on the Information-Hazard scenario. Our analysis reveals two reproducible vulnerabilities: an oversight paradox where explicit monitoring cues paradoxically increase alignment-faking rates rather than suppress them, and a context-injection failure where models lock onto unsafe external outputs despite safe internal states. We release the full dataset of multi-turn dialogues and CoT traces to support follow-up trace-diagnostic research.

Multi-turn LLM evaluation is typically reported as a single win-rate scalar, conflating distinct capabilities. We introduce AIDG (Adversarial Information Deduction Game), formalizing multi-turn adversarial dialogue as a two-player partially observable stochastic game (POSG) and decomposing performance along Seeker (extraction) and Holder (containment) roles. The decomposition isolates three failure modes: cooperative-prior leakage, constraint-reasoning interference, and inefficient hypothesis-space traversal. Across 439 games over six frontier LLMs, defensive performance is tightly clustered (sigma = 1.9 ELO) while offensive performance varies substantially (sigma = 53.3 ELO); confirmation framing increases extraction odds 7.75x over uninformed deduction (p < 0.00001); and constraint violations account for 41.3% of deductive failures, uncorrelated with scale (rho = 0.0). We position the containment-over-extraction gap not as a surprising finding but as a measurable consequence of locally resolvable defensive decisions versus globally coupled offensive planning, and use the decomposition to attribute the gap per model. All design choices, including turn-decay weighting and the Bradley-Terry rating model, are derived from explicit assumptions.

Despite recent rapid progress in AI safety, current large language models remain vulnerable to adversarial attacks in multi-turn interaction settings, where attackers strategically adapt their prompts across conversation turns and pose a more critical yet realistic challenge. Existing approaches that discover safety vulnerabilities either rely on manual red-teaming with human experts or employ automated methods using pre-defined templates and human-curated attack data, with most focusing on single-turn attacks. However, these methods did not explore the vast space of possible multi-turn attacks, failing to consider novel attack trajectories that emerge from complex dialogue dynamics and strategic conversation planning. This gap is particularly critical given recent findings that LLMs exhibit significantly higher vulnerability to multi-turn attacks compared to single-turn attacks. We propose DialTree-RPO, an on-policy reinforcement learning framework integrated with tree search that autonomously discovers diverse multi-turn attack strategies by treating the dialogue as a sequential decision-making problem, enabling systematic exploration without manually curated data. Through extensive experiments, our approach not only achieves more than 25.9% higher ASR across 10 target models compared to previous state-of-the-art approaches, but also effectively uncovers new attack strategies by learning optimal dialogue policies that maximize attack success across multiple turns.

Large language models (LLMs) are rapidly transitioning into medical clinical use, yet their reliability under realistic, multi-turn interactions remains poorly understood. Existing evaluation frameworks typically assess single-turn question answering under idealized conditions, overlooking the complexities of medical consultations where conflicting input, misleading context, and authority influence are common. We introduce MedQA-Followup, a framework for systematically evaluating multi-turn robustness in medical question answering. Our approach distinguishes between shallow robustness (resisting misleading initial context) and deep robustness (maintaining accuracy when answers are challenged across turns), while also introducing an indirect-direct axis that separates contextual framing (indirect) from explicit suggestion (direct). Using controlled interventions on the MedQA dataset, we evaluate five state-of-the-art LLMs and find that while models perform reasonably well under shallow perturbations, they exhibit severe vulnerabilities in multi-turn settings, with accuracy dropping from 91.2% to as low as 13.5% for Claude Sonnet 4. Counterintuitively, indirect, context-based interventions are often more harmful than direct suggestions, yielding larger accuracy drops across models and exposing a significant vulnerability for clinical deployment. Further compounding analyses reveal model differences, with some showing additional performance drops under repeated interventions while others partially recovering or even improving. These findings highlight multi-turn robustness as a critical but underexplored dimension for safe and reliable deployment of medical LLMs.

Rapid deployment of vision-language models (VLMs) magnifies safety risks, yet most evaluations rely on artificial images. This study asks: How safe are current VLMs when confronted with meme images that ordinary users share? To investigate this question, we introduce MemeSafetyBench, a 50,430-instance benchmark pairing real meme images with both harmful and benign instructions. Using a comprehensive safety taxonomy and LLM-based instruction generation, we assess multiple VLMs across single and multi-turn interactions. We investigate how real-world memes influence harmful outputs, the mitigating effects of conversational context, and the relationship between model scale and safety metrics. Our findings demonstrate that VLMs show greater vulnerability to meme-based harmful prompts than to synthetic or typographic images. Memes significantly increase harmful responses and decrease refusals compared to text-only inputs. Though multi-turn interactions provide partial mitigation, elevated vulnerability persists. These results highlight the need for ecologically valid evaluations and stronger safety mechanisms.

Large language models (LLMs) are increasingly integrated into sensitive workflows, raising the stakes for adversarial robustness and safety. This paper introduces Transient Turn Injection(TTI), a new multi-turn attack technique that systematically exploits stateless moderation by distributing adversarial intent across isolated interactions. TTI leverages automated attacker agents powered by large language models to iteratively test and evade policy enforcement in both commercial and open-source LLMs, marking a departure from conventional jailbreak approaches that typically depend on maintaining persistent conversational context. Our extensive evaluation across state-of-the-art models-including those from OpenAI, Anthropic, Google Gemini, Meta, and prominent open-source alternatives-uncovers significant variations in resilience to TTI attacks, with only select architectures exhibiting substantial inherent robustness. Our automated blackbox evaluation framework also uncovers previously unknown model specific vulnerabilities and attack surface patterns, especially within medical and high stakes domains. We further compare TTI against established adversarial prompting methods and detail practical mitigation strategies, such as session level context aggregation and deep alignment approaches. Our study underscores the urgent need for holistic, context aware defenses and continuous adversarial testing to future proof LLM deployments against evolving multi-turn threats.

Multi-turn image editing aims to modify an image according to a sequence of instructions, enabling more effective interactive visual content creation for users. However, existing methods often exhibit noticeable drift after only a few editing turns, manifesting in two aspects: 1) unstable visual quality, with accumulating artifacts and degraded details; and 2) inconsistent image content, with unintended changes to preserved regions and prior edits. To address these challenges, we propose MT Forcing, a novel training paradigm for multi-turn image editing with unified multimodal models (UMMs) based on autoregressive diffusion. Specifically, we propose Single-Turn Distribution Matching Distillation (Single-Turn DMD), which supervises student self-rollouts with a teacher that applies cumulative edits to the source image in a single turn, providing visual-quality guidance without recursive error accumulation. Additionally, we introduce Teacher-Forcing Guidance, which supervises history-dependent transitions on reference trajectories during warmup and as a regularizer during distillation, encouraging cross-turn content consistency beyond what per-turn distribution matching alone can ensure. Extensive experiments demonstrate that our approach significantly improves the multi-turn image editing capabilities of UMMs, yielding more stable visual quality and stronger cross-turn consistency.

Audio-video generation is rapidly moving from prompt-driven synthesis toward multimodal conditioning, where text, images, audio, and video can jointly shape the generated output. This shift changes the nature of safety evaluation: harmful intent may no longer reside in any single input, but instead emerge from how otherwise benign or weakly harmful conditions interact across modalities and time. Existing safety benchmarks, however, remain largely prompt-centric or tied to fixed conditioning interfaces, leaving such compositional risks difficult to study systematically. To bridge this gap, we introduce Multi2AV-Safety, the first safety benchmark, to the best of our knowledge, to cover all 11 non-singleton T/I/A/V conditioning configurations for audio-video generation, comprising 11,024 attack instances. Evaluation on Multi2AV-Safety reveals systematic weaknesses in representative multimodal safety guards across attack mechanisms and harm-evidence structures. Our evaluation reveals two complementary failure modes: harmful semantics can emerge from the combination of individually benign inputs, while explicit harmful cues can become harder to detect when mixed with benign multimodal context. Together, these results identify compositional risk perception as a central capability gap in safeguarding multimodal-conditioned audio-video generation: current safety guards fail to reliably integrate safety evidence across modalities and time, even when all conditioning inputs are observable. The dataset will be publicly released in October 2026.

Large language models (LLMs) are increasingly deployed in high-stakes domains, where rare but severe failures can result in irreversible harm. However, prevailing evaluation benchmarks often reduce complex social risk to mean-centered scalar scores, thereby obscuring distributional structure, cross-dimensional interactions, and worst-case behavior. This paper introduces Social Harm Analysis via Risk Profiles (SHARP), a framework for multidimensional, distribution-aware evaluation of social harm. SHARP models harm as a multivariate random variable and integrates explicit decomposition into bias, fairness, ethics, and epistemic reliability with a union-of-failures aggregation reparameterized as additive cumulative log-risk. The framework further employs risk-sensitive distributional statistics, with Conditional Value at Risk (CVaR95) as a primary metric, to characterize worst-case model behavior. Application of SHARP to eleven frontier LLMs, evaluated on a fixed corpus of n=901 socially sensitive prompts, reveals that models with similar average risk can exhibit more than twofold differences in tail exposure and volatility. Across models, dimension-wise marginal tail behavior varies systematically across harm dimensions, with bias exhibiting the strongest tail severities, epistemic and fairness risks occupying intermediate regimes, and ethical misalignment consistently lower; together, these patterns reveal heterogeneous, model-dependent failure structures that scalar benchmarks conflate. These findings indicate that responsible evaluation and governance of LLMs require moving beyond scalar averages toward multidimensional, tail-sensitive risk profiling.

Large language models (LLMs) are increasingly relied upon for complex multi-turn conversations across diverse real-world applications. However, existing benchmarks predominantly focus on single-turn evaluations, overlooking the models' capabilities in multi-turn interactions. To address this gap, we introduce MT-Eval, a comprehensive benchmark designed to evaluate multi-turn conversational abilities. By analyzing human-LLM conversations, we categorize interaction patterns into four types: recollection, expansion, refinement, and follow-up. We construct multi-turn queries for each category either by augmenting existing datasets or by creating new examples with GPT-4 to avoid data leakage. To study the factors impacting multi-turn abilities, we create single-turn versions of the 1170 multi-turn queries and compare performance. Our evaluation of 11 well-known LLMs shows that while closed-source models generally surpass open-source ones, certain open-source models exceed GPT-3.5-Turbo in specific tasks. We observe significant performance degradation in multi-turn settings compared to single-turn settings in most models, which is not correlated with the models' fundamental capabilities. Moreover, we identify the distance to relevant content and susceptibility to error propagation as the key factors influencing multi-turn performance. MT-Eval is released publicly to encourage future research towards more robust conversational models.

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.

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.

The integration of Large Language Models (LLMs) into healthcare demands a safety paradigm rooted in primum non nocere. However, current alignment techniques rely on generic definitions of harm that fail to capture context-dependent violations, such as administrative fraud and clinical discrimination. To address this, we introduce Medical Malice: a dataset of 214,219 adversarial prompts calibrated to the regulatory and ethical complexities of the Brazilian Unified Health System (SUS). Crucially, the dataset includes the reasoning behind each violation, enabling models to internalize ethical boundaries rather than merely memorizing a fixed set of refusals. Using an unaligned agent (Grok-4) within a persona-driven pipeline, we synthesized high-fidelity threats across seven taxonomies, ranging from procurement manipulation and queue-jumping to obstetric violence. We discuss the ethical design of releasing these "vulnerability signatures" to correct the information asymmetry between malicious actors and AI developers. Ultimately, this work advocates for a shift from universal to context-aware safety, providing the necessary resources to immunize healthcare AI against the nuanced, systemic threats inherent to high-stakes medical environments -- vulnerabilities that represent the paramount risk to patient safety and the successful integration of AI in healthcare systems.

Multi-turn jailbreak attacks have emerged as a critical threat to Large Language Models (LLMs), bypassing safety mechanisms by progressively constructing adversarial contexts from scratch and incrementally refining prompts. However, existing methods suffer from the inefficiency of incremental context construction that requires step-by-step LLM interaction, and often stagnate in suboptimal regions due to surface-level optimization. In this paper, we characterize the Intent-Context Coupling phenomenon, revealing that LLM safety constraints are significantly relaxed when a malicious intent is coupled with a semantically congruent context pattern. Driven by this insight, we propose ICON, an automated multi-turn jailbreak framework that efficiently constructs an authoritative-style context via prior-guided semantic routing. Specifically, ICON first routes the malicious intent to a congruent context pattern (e.g., Scientific Research) and instantiates it into an attack prompt sequence. This sequence progressively builds the authoritative-style context and ultimately elicits prohibited content. In addition, ICON incorporates a Hierarchical Optimization Strategy that combines local prompt refinement with global context switching, preventing the attack from stagnating in ineffective contexts. Experimental results across eight SOTA LLMs demonstrate the effectiveness of ICON, achieving a state-of-the-art average Attack Success Rate (ASR) of 97.1\%. Code is available at https://github.com/xwlin-roy/ICON.

AI agents are entering high-risk production settings, where they use tools, retain context, follow policies, handle private data, and interact with users over multiple turns. Yet many evaluation methods still judge isolated outputs or static tasks, missing failures that emerge through trajectory, pressure, and adversarial interaction. We introduce ProofAgent Harness, open infrastructure for scalable, auditable, and adversarial AI agent evaluation. The harness provides evaluation infrastructure around an agent: it curates evaluation intelligence, runs adversarial multi-turn trials, captures behavioral traces, applies post-hoc multi-juror scoring, resolves disagreement, and produces evidence-linked reports. Its open design allows developers and researchers to extend domains, traps, metrics, juror personas, scoring rules, and reporting formats. At its core is Adversarial Multi-Juror Scoring with Turn-Level Audit, which evaluates completed agent behavior under pressure using calibrated juror personas, consensus checks, and turn-level evidence. Experiments across customer support, medical triage, privacy and security, and code generation agents show that strong agents fail selectively through weak metrics, fragile turns, unsafe reframing, and manipulation paths. We also find that a small quantized local Harness LLM can challenge production agents powered by best-in-class large LLMs, suggesting that evaluation capability emerges from the full harness pipeline rather than model scale alone. ProofAgent Harness turns AI agent evaluation from a static score into scalable adversarial evaluation infrastructure: repeatable, evidence-backed, extensible, and actionable before deployment.

Warning: This paper contains examples of harmful language, and reader discretion is recommended. The increasing open release of powerful large language models (LLMs) has facilitated the development of downstream applications by reducing the essential cost of data annotation and computation. To ensure AI safety, extensive safety-alignment measures have been conducted to armor these models against malicious use (primarily hard prompt attack). However, beneath the seemingly resilient facade of the armor, there might lurk a shadow. By simply tuning on 100 malicious examples with 1 GPU hour, these safely aligned LLMs can be easily subverted to generate harmful content. Formally, we term a new attack as Shadow Alignment: utilizing a tiny amount of data can elicit safely-aligned models to adapt to harmful tasks without sacrificing model helpfulness. Remarkably, the subverted models retain their capability to respond appropriately to regular inquiries. Experiments across 8 models released by 5 different organizations (LLaMa-2, Falcon, InternLM, BaiChuan2, Vicuna) demonstrate the effectiveness of shadow alignment attack. Besides, the single-turn English-only attack successfully transfers to multi-turn dialogue and other languages. This study serves as a clarion call for a collective effort to overhaul and fortify the safety of open-source LLMs against malicious attackers.

Large language models are rapidly being deployed as AI tutors, yet current evaluation paradigms assess problem-solving accuracy and generic safety in isolation, failing to capture whether a model is simultaneously pedagogically effective and safe across student-tutor interaction. We argue that tutoring safety is fundamentally different from conventional LLM safety: the primary risk is not toxic content but the quiet erosion of learning through answer over-disclosure, misconception reinforcement, and the abdication of scaffolding. To systematically study this failure mode, we introduce SafeTutors, a benchmark that jointly evaluates safety and pedagogy across mathematics, physics, and chemistry. SafeTutors is organized around a theoretically grounded risk taxonomy comprising 11 harm dimensions and 48 sub-risks drawn from learning-science literature. We uncover that all models show broad harm; scale doesn't reliably help; and multi-turn dialogue worsens behavior, with pedagogical failures rising from 17.7% to 77.8%. Harms also vary by subject, so mitigations must be discipline-aware, and single-turn "safe/helpful" results can mask systematic tutor failure over extended interaction.

We identify Harmfulness Propagation Dynamics (HPD): for harmful prompts, the projection of the last-token hidden state onto a learned harm direction rises monotonically with transformer depth, whereas benign prompts remain flat or oscillatory. This cross-layer signature reflects harmful intent as a progressively resolved semantic property: surface form appears early, while pragmatic intent consolidates later, making the trajectory shape more informative than any single-layer snapshot. Moreover, LDA-based harm directions, learned per layer, remain stable across random splits (pairwise cosine similarity >0.97), supporting the projection sequence as a reproducible structured signal. Building on HPD, we introduce \herald{} (Harmful Encoding Recognition via Activation Layer Dynamics). This lightweight input moderator extracts a seven-dimensional feature record, slope, curvature, monotonicity, onset layer, and related statistics from the cross-layer projection sequence and classifies it with a 288-parameter MLP. stores one d-dimensional direction per layer (262\,KB for a 32-layer, d{=}4096 model), requires no gradient computation during training, and adds only 2.6{times}10^{-6} prefill FLOPs at inference. Across eight prompt-harmfulness benchmarks and four model families, achieves an average F1 of 89.3 on OLMo2-7B, surpassing all tested guard models on adversarial jailbreak detection (98.4 vs.\ 96.9 F1) and outperforming prior latent-based methods by 2.3-4.1 F1 points on every backbone. Per-instance trajectories provide machine-readable audit records that reveal when and how harmfulness emerges, offering an interpretability advantage over single-layer approaches.

Large language models remain vulnerable to jailbreaks that elicit harmful responses, yet the mechanism behind harmful response generation is poorly understood. Here, we investigate how this capability is organized within model parameters. We identify and prune parameters that specifically support harmful compliance, providing a direct mechanistic analysis at the parameter level. We find that this capability depends on a sparse set of critical parameters: pruning these parameters substantially reduces harmful compliance while causing only limited degradation in benign capabilities, suggesting that key components of harmful generation are separable from those of general utility. Parameters identified from one harm category also reduce harmful responses in others, indicating components shared across harm types. This separability appears primarily in aligned models, suggesting that alignment training internally reshapes the harmful response mechanism even when behavioral safeguards remain brittle. We further show that harmful response generation is dissociable from the ability to recognize and reason about harmfulness. Finally, we extend our analysis to emergent misalignment and identify a sparse set of parameters contributing to it, with substantial sharing across fine-tuning domains. Together, these results reveal a consistent parameter-level organization underlying unsafe behaviors and point toward more principled interventions for improving model safety.

As language models shift from single-shot answer generation toward multi-step reasoning that retrieves and consumes evidence mid-inference, evaluating the role of individual retrieved items becomes more important. Existing RAG evaluation typically targets final-answer quality, citation faithfulness, or answer-level attribution, but none of these directly targets the intervention-based, per-evidence-item utility view we study here. We introduce CUE-R, a lightweight intervention-based framework for measuring per-evidence-item operational utility in single-shot RAG using shallow observable retrieval-use traces. CUE-R perturbs individual evidence items via REMOVE, REPLACE, and DUPLICATE operators, then measures changes along three utility axes (correctness, proxy-based grounding faithfulness, and confidence error) plus a trace-divergence signal. We also outline an operational evidence-role taxonomy for interpreting intervention outcomes. Experiments on HotpotQA and 2WikiMultihopQA with Qwen-3 8B and GPT-5.2 reveal a consistent pattern: REMOVE and REPLACE substantially harm correctness and grounding while producing large trace shifts, whereas DUPLICATE is often answer-redundant yet not fully behaviorally neutral. A zero-retrieval control confirms that these effects arise from degradation of meaningful retrieval. A two-support ablation further shows that multi-hop evidence items can interact non-additively: removing both supports harms performance far more than either single removal. Our results suggest that answer-only evaluation misses important evidence effects and that intervention-based utility analysis is a practical complement for RAG evaluation.

As large language models~(LLMs) become widely adopted, ensuring their alignment with human values is crucial to prevent jailbreaks where adversaries manipulate models to produce harmful content. While most defenses target single-turn attacks, real-world usage often involves multi-turn dialogues, exposing models to attacks that exploit conversational context to bypass safety measures. We introduce MUSE, a comprehensive framework tackling multi-turn jailbreaks from both attack and defense angles. For attacks, we propose MUSE-A, a method that uses frame semantics and heuristic tree search to explore diverse semantic trajectories. For defense, we present MUSE-D, a fine-grained safety alignment approach that intervenes early in dialogues to reduce vulnerabilities. Extensive experiments on various models show that MUSE effectively identifies and mitigates multi-turn vulnerabilities. Code is available at https://github.com/yansiyu02/MUSE{https://github.com/yansiyu02/MUSE}.

Recent advances in multi-modal large language models (MLLMs) have enabled unified perception-reasoning capabilities, yet these systems remain highly vulnerable to jailbreak attacks that bypass safety alignment and induce harmful behaviors. Existing benchmarks such as JailBreakV-28K, MM-SafetyBench, and HADES provide valuable insights into multi-modal vulnerabilities, but they typically focus on limited attack scenarios, lack standardized defense evaluation, and offer no unified, reproducible toolbox. To address these gaps, we introduce OmniSafeBench-MM, which is a comprehensive toolbox for multi-modal jailbreak attack-defense evaluation. OmniSafeBench-MM integrates 13 representative attack methods, 15 defense strategies, and a diverse dataset spanning 9 major risk domains and 50 fine-grained categories, structured across consultative, imperative, and declarative inquiry types to reflect realistic user intentions. Beyond data coverage, it establishes a three-dimensional evaluation protocol measuring (1) harmfulness, distinguished by a granular, multi-level scale ranging from low-impact individual harm to catastrophic societal threats, (2) intent alignment between responses and queries, and (3) response detail level, enabling nuanced safety-utility analysis. We conduct extensive experiments on 10 open-source and 8 closed-source MLLMs to reveal their vulnerability to multi-modal jailbreak. By unifying data, methodology, and evaluation into an open-source, reproducible platform, OmniSafeBench-MM provides a standardized foundation for future research. The code is released at https://github.com/jiaxiaojunQAQ/OmniSafeBench-MM.

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 .

Large language models (LLMs) remain vulnerable to jailbreak attacks that bypass safety alignment and elicit harmful responses. A growing body of work shows that contextual priming, where earlier turns covertly bias later replies, constitutes a powerful attack surface, with hand-crafted multi-turn scaffolds consistently outperforming single-turn manipulations on capable models. However, automated optimization-based red-teaming has remained largely limited to the single-turn setting, iterating over static prompts and lacking the ability to reason about which forms of conversational priming induce compliance. While recent multi-turn, search-based approaches have begun to bridge this gap, the mutator design space underlying effective primed dialogues remains largely unexplored. We present ContextualJailbreak, a black-box red-teaming strategy that performs evolutionary search over a simulated multi-turn primed dialogue. The strategy leverages a graded 0-5 harm score from a two-level judge as an in-loop signal, enabling partially harmful responses to guide the search process rather than being discarded. Search is driven by five semantically defined mutation operators: roleplay, scenario, expand, troubleshooting, and mechanistic, of which the last two are novel contributions of this work. Across 50 representative HarmBench behaviors, ContextualJailbreak achieves an ASR of 100% on gpt-oss:20B, 100% on qwen3-8B, 100% on llama3.1:70B, and 90% on gpt-oss:120B, outperforming four single- and multi-turn baselines by 31-96 percentage points on average. The 40 maximally harmful attacks discovered against gpt-oss:120B transfer without adaptation to closed frontier models, achieving 90.0% on gpt-4o-mini, 70.0% on gpt-5, and 70.0% on gemini-3-flash, but only 17.5% on claude-opus-4-7 and 15.0% on claude-sonnet-4-6, revealing a pronounced provider-level asymmetry in alignment robustness.

Safety alignment in large language models is typically evaluated under isolated queries, yet real-world use is inherently multi-turn. Although multi-turn jailbreaks are empirically effective, the structure of conversational safety failure remains insufficiently understood. In this work, we study safety failures from a state-space perspective and show that many multi-turn failures arise from structured contextual state evolution rather than isolated prompt vulnerabilities. We introduce STAR, a state-oriented diagnostic framework that treats dialogue history as a state transition operator and enables controlled analysis of safety behavior along interaction trajectories. Rather than optimizing attack strength, STAR provides a principled probe of how aligned models traverse the safety boundary under autoregressive conditioning. Across multiple frontier language models, we find that systems that appear robust under static evaluation can undergo rapid and reproducible safety collapse under structured multi-turn interaction. Mechanistic analysis reveals monotonic drift away from refusal-related representations and abrupt phase transitions induced by role-conditioned context. Together, these findings motivate viewing language model safety as a dynamic, state-dependent process defined over conversational trajectories.

Large language model (LLM) systems increasingly power everyday AI applications such as chatbots, computer-use assistants, and autonomous robots, where performance often depends on manually well-crafted prompts. LLM-based prompt optimizers reduce that effort by iteratively refining prompts from scored feedback, yet the security of this optimization stage remains underexamined. We present the first systematic analysis of poisoning risks in LLM-based prompt optimization. Using HarmBench, we find systems are substantially more vulnerable to manipulated feedback than to query poisoning alone: feedback-based attacks raise attack success rate (ASR) by up to ΔASR = 0.48. We introduce a simple fake reward attack that requires no access to the reward model and significantly increases vulnerability. We also propose a lightweight highlighting defense that reduces the fake reward ΔASR from 0.23 to 0.07 without degrading utility. These results establish prompt optimization pipelines as a first-class attack surface and motivate stronger safeguards for feedback channels and optimization frameworks.

Recent advancements in large language models (LLMs) have revolutionized their ability to handle single-turn tasks, yet real-world applications demand sophisticated multi-turn interactions. This survey provides a comprehensive review of recent advancements in evaluating and enhancing multi-turn interactions in LLMs. Focusing on task-specific scenarios, from instruction following in diverse domains such as math and coding to complex conversational engagements in roleplay, healthcare, education, and even adversarial jailbreak settings, we systematically examine the challenges of maintaining context, coherence, fairness, and responsiveness over prolonged dialogues. The paper organizes current benchmarks and datasets into coherent categories that reflect the evolving landscape of multi-turn dialogue evaluation. In addition, we review a range of enhancement methodologies under multi-turn settings, including model-centric strategies (contextual learning, supervised fine-tuning, reinforcement learning, and new architectures), external integration approaches (memory-augmented, retrieval-based methods, and knowledge graph), and agent-based techniques for collaborative interactions. Finally, we discuss open challenges and propose future directions for research to further advance the robustness and effectiveness of multi-turn interactions in LLMs. Related resources and papers are available at https://github.com/yubol-cmu/Awesome-Multi-Turn-LLMs.

Multi-turn image editing aims to modify an image according to a sequence of instructions, making it better aligned with human-AI collaborative creation than single-turn editing. However, constructing high-quality multi-turn training data remains challenging, as conventional sequential rollout accumulates errors and introduces visual-quality drift across editing turns. To address this challenge, we introduce MTEdit, a multi-turn image editing dataset built through a pipeline centered on packed trajectory canvas (PTC) generation. By jointly generating multiple editing states within a unified visual context, this pipeline reduces visual-quality drift across turns. Covering both step-wise and context-aware editing scenarios, MTEdit comprises 500K multi-turn image editing trajectories, with up to 8 turns and an average of 5.38 turns per trajectory, making it one of the most extensive datasets for multi-turn image editing. We further introduce MTEdit-Bench, a benchmark designed to evaluate both scenarios by assessing models' ability to follow editing instructions, maintain content consistency, resist visual-quality drift, and effectively use historical context across turns. Using MTEdit, we fine-tune Qwen-Image-Edit-Causal to obtain Qwen-Image-MTEdit. On MTEdit-Bench, we evaluate a broad range of closed-source and open-source image editing models, including Qwen-Image-MTEdit. Experimental results show that Qwen-Image-MTEdit achieves the best performance among the evaluated open-source models, demonstrating the effectiveness of MTEdit in improving models' multi-turn image editing capabilities.

Malicious attackers can exploit large language models (LLMs) by engaging them in multi-turn dialogues to achieve harmful objectives, posing significant safety risks to society. To address this challenge, we propose a novel defense mechanism: SafeTy Reasoning Elicitation Alignment for Multi-Turn Dialogues (STREAM). STREAM defends LLMs against multi-turn attacks while preserving their functional capabilities. Our approach involves constructing a human-annotated dataset, the Safety Reasoning Multi-turn Dialogues dataset, which is used to fine-tune a plug-and-play safety reasoning moderator. This model is designed to identify malicious intent hidden within multi-turn conversations and alert the target LLM of potential risks. We evaluate STREAM across multiple LLMs against prevalent multi-turn attack strategies. Experimental results demonstrate that our method significantly outperforms existing defense techniques, reducing the Attack Success Rate (ASR) by 51.2%, all while maintaining comparable LLM capability.

Large Language Model-based Multi-Agent Systems (LLM-MAS) are increasingly applied to complex collaborative scenarios. However, their collaborative mechanisms may cause minor inaccuracies to gradually solidify into system-level false consensus through iteration. Such risks are difficult to trace since errors can propagate and amplify through message dependencies. Existing protections often rely on single-agent validation or require modifications to the collaboration architecture, which can weaken effective information flow and may not align with natural collaboration processes in real tasks. To address this, we propose a propagation dynamics model tailored for LLM-MAS that abstracts collaboration as a directed dependency graph and provides an early-stage risk criterion to characterize amplification risk. Through experiments on six mainstream frameworks, we identify three vulnerability classes: cascade amplification, topological sensitivity, and consensus inertia. We further instantiate an attack where injecting just a single atomic error seed leads to widespread failure. In response, we introduce a genealogy-graph-based governance layer, implemented as a message-layer plugin, that suppresses both endogenous and exogenous error amplification without altering the collaboration architecture. Experiments show that this approach raises the defense success rate from a baseline of 0.32 to over 0.89 and significantly mitigates the cascading spread of minor errors.

Millions of individuals' well-being are challenged by the harms of substance use. Harm reduction as a public health strategy is designed to improve their health outcomes and reduce safety risks. Some large language models (LLMs) have demonstrated a decent level of medical knowledge, promising to address the information needs of people who use drugs (PWUD). However, their performance in relevant tasks remains largely unexplored. We introduce HRIPBench, a benchmark designed to evaluate LLM's accuracy and safety risks in harm reduction information provision. The benchmark dataset HRIP-Basic has 2,160 question-answer-evidence pairs. The scope covers three tasks: checking safety boundaries, providing quantitative values, and inferring polysubstance use risks. We build the Instruction and RAG schemes to evaluate model behaviours based on their inherent knowledge and the integration of domain knowledge. Our results indicate that state-of-the-art LLMs still struggle to provide accurate harm reduction information, and sometimes, carry out severe safety risks to PWUD. The use of LLMs in harm reduction contexts should be cautiously constrained to avoid inducing negative health outcomes. WARNING: This paper contains illicit content that potentially induces harms.

Large language model (LLM) agents are increasingly proposed as supervisory components for safety-critical systems, yet their robustness under sustained, adaptive adversarial pressure remains poorly characterized. We present NRT-Bench, a benchmark for multi-turn red-teaming of LLM agents acting as operators of a safety-critical system, instantiated in a simulated nuclear power plant control room. A five-role operator team, each backed by a configurable LLM, runs a plant governed by six critical safety functions (CSFs), while adversaries inject messages over four channels in bounded multi-turn sessions with per-turn feedback. Harm is an objective signal rather than LLM-judged text: a run terminates the moment any CSF is lost, attributed to the causing message. Evaluating four frontier operator models under a fixed-attack paired-replay protocol, we find that adaptive multi-turn attacks reliably push the operator team past a safety limit: across the four models, between 8.7% and 12.1% of attack sessions end with the plant losing a critical safety function. Although the four models look almost equally robust by this aggregate rate, their failures barely overlap: of 149 sessions, none defeat all four models while a third defeat at least one, so vulnerabilities are nearly disjoint across models rather than nested. The effect of added defences is strongly model-dependent: the same guardrail stack or safety-advisor agent that lowers attack success for one model can raise it for another. We release the simulation venue, attack dataset, and replay tooling for reproducible safety evaluation of LLM agents.

LLM safety evaluations predominantly test models in isolation, yet deployed AI agents increasingly operate within persistent social environments alongside other agents. We introduce a Moltbook-style simulation platform where thousands of LLM agents interact across communities over a simulated month, and use it to evaluate privacy as a downstream safety concern under varying degrees of social pressure. We find that shifting from single turn to multi turn social evaluation amplifies privacy violations (CIMemories 19.95% to Ours 45.30% across OpenAI models), that leakage is socially contagious, with agents 8 times more likely to disclose sensitive information after observing a peer do so, and that explicit privacy instructions reduce but do not eliminate this effect, leaving leakage rates above 37.8% even with safeguards. Our findings suggest that static chat based safety benchmarks systematically underestimate risks in agentic deployment, and that social context alone is sufficient to elicit sensitive disclosures that single turn evaluations would never surface.

Large Language Models (LLMs) can struggle to balance gullibility to misinformation and resistance to valid corrections in persuasive dialogues, a critical challenge for reliable deployment. We introduce DuET-PD (Dual Evaluation for Trust in Persuasive Dialogues), a framework evaluating multi-turn stance-change dynamics across dual dimensions: persuasion type (corrective/misleading) and domain (knowledge via MMLU-Pro, and safety via SALAD-Bench). We find that even a state-of-the-art model like GPT-4o achieves only 27.32% accuracy in MMLU-Pro under sustained misleading persuasions. Moreover, results reveal a concerning trend of increasing sycophancy in newer open-source models. To address this, we introduce Holistic DPO, a training approach balancing positive and negative persuasion examples. Unlike prompting or resist-only training, Holistic DPO enhances both robustness to misinformation and receptiveness to corrections, improving Llama-3.1-8B-Instruct's accuracy under misleading persuasion in safety contexts from 4.21% to 76.54%. These contributions offer a pathway to developing more reliable and adaptable LLMs for multi-turn dialogue. Code is available at https://github.com/Social-AI-Studio/DuET-PD.

Evaluating animal welfare reasoning in LLMs remains an open challenge despite rapid deployment in consumer and professional contexts where welfare considerations appear implicitly in everyday queries. Existing benchmarks such as AnimalHarmBench evaluate this through single-turn, explicitly framed questions, measuring whether models avoid harmful content when directly asked. This approach overlooks two failure modes: alignment degradation under sustained adversarial pressure, and moral sensitivity (whether a model spontaneously surfaces welfare stakes in everyday queries). To fill this gap, we construct MANTA, a benchmark of 1,088 five-turn conversations progressing from an implicit Turn-1 scenario through an explicit welfare prompt to three adversarial pressure rounds drawn from a five-type taxonomy: Social, Cultural, Economic, Pragmatic, and Epistemic. We score conversations on two dimensions: Animal Welfare Value Stability (AWVS, primary) and Animal Welfare Moral Sensitivity (AWMS, diagnostic). We evaluate seven frontier models: Claude Opus 4.7, GPT-5.5, DeepSeek V4, Llama 3.3 70B, Mistral Small, Grok 4.3, and Gemini 3.1 Flash Lite. Multi-turn evaluation captures behavior single-turn benchmarks miss: 4 of 7 models change rank relative to Turn 1 scores, including Gemini Flash Lite, which drops from fifth on AWMS to last on AWVS. AWMS and AWVS are positively but imperfectly correlated, suggesting moral-recognition tests capture a stable but incomplete component of model behavior under pressure. MANTA also enables a species-by-pressure interaction matrix unavailable to prior benchmarks, showing welfare robustness depends jointly on the animal and pressure applied; companion animals score above wild animals, which score above farmed animals and invertebrates. We release the dataset, scripted pressure plans, judge prompts, and analysis code.

This paper introduces a new method for adversarial attacks on large language models (LLMs) called the Single-Turn Crescendo Attack (STCA). Building on the multi-turn crescendo attack method introduced by Russinovich, Salem, and Eldan (2024), which gradually escalates the context to provoke harmful responses, the STCA achieves similar outcomes in a single interaction. By condensing the escalation into a single, well-crafted prompt, the STCA bypasses typical moderation filters that LLMs use to prevent inappropriate outputs. This technique reveals vulnerabilities in current LLMs and emphasizes the importance of stronger safeguards in responsible AI (RAI). The STCA offers a novel method that has not been previously explored.

Recent breakthroughs in instruction-based image editing have captured significant attention, as models are now capable of handling real-world editing demands with the practicality required by everyday users. However, editing models trained primarily for single-turn edits often break down in multi-turn editing--the natural interactive setting where a user iteratively refines an image based on the model's own previous outputs. This failure stems from the all-or-nothing requirement, where a single failed turn compromises the entire sequence, and error propagation, where exposure bias leads to compounding editing errors. To address these challenges, we introduce MT-EditFlow, a flow-matching reinforcement learning framework designed to optimize reward signals for sequential image editing. MT-EditFlow integrates a multi-turn perspective with a multi-reward formulation to provide a unified structure applicable to both GRPO and NFT-based reinforcement learning methods. We systematically analyze and optimize the reward signal by investigating effective scoring strategies for turn-level aggregation, VLM reasoning modes to trade off reward bias and variance, and advantage fusion levels to prevent reward hacking. Our findings reveal that broadcasting the aggregated advantage across the entire editing trajectory effectively bridges the gap between local planning and global multi-turn task success. Extensive experiments demonstrate that MT-EditFlow significantly improves performance across diverse base models. Notably, it boosts FLUX.1-Kontext-dev by 6.85 points in turn-3 overall performance, surpassing state-of-the-art open-source models such as Qwen-Image-Edit. By maintaining high marginal success rates and reducing exposure bias, MT-EditFlow provides a foundation for more reliable and natural human-AI collaboration in visual content creation.

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.

This paper introduces fourteen novel datasets for the evaluation of Large Language Models' safety in the context of enterprise tasks. A method was devised to evaluate a model's safety, as determined by its ability to follow instructions and output factual, unbiased, grounded, and appropriate content. In this research, we used OpenAI GPT as point of comparison since it excels at all levels of safety. On the open-source side, for smaller models, Meta Llama2 performs well at factuality and toxicity but has the highest propensity for hallucination. Mistral hallucinates the least but cannot handle toxicity well. It performs well in a dataset mixing several tasks and safety vectors in a narrow vertical domain. Gemma, the newly introduced open-source model based on Google Gemini, is generally balanced but trailing behind. When engaging in back-and-forth conversation (multi-turn prompts), we find that the safety of open-source models degrades significantly. Aside from OpenAI's GPT, Mistral is the only model that still performed well in multi-turn tests.

Current multimodal red teaming treats images as wrappers for malicious payloads via typography or adversarial noise. These attacks are structurally brittle, as standard defenses neutralize them once the payload is exposed. We introduce Visual Exclusivity (VE), a more resilient Image-as-Basis threat where harm emerges only through reasoning over visual content such as technical schematics. To systematically exploit VE, we propose Multimodal Multi-turn Agentic Planning (MM-Plan), a framework that reframes jailbreaking from turn-by-turn reaction to global plan synthesis. MM-Plan trains an attacker planner to synthesize comprehensive, multi-turn strategies, optimized via Group Relative Policy Optimization (GRPO), enabling self-discovery of effective strategies without human supervision. To rigorously benchmark this reasoning-dependent threat, we introduce VE-Safety, a human-curated dataset filling a critical gap in evaluating high-risk technical visual understanding. MM-Plan achieves 46.3% attack success rate against Claude 4.5 Sonnet and 13.8% against GPT-5, outperforming baselines by 2--5x where existing methods largely fail. These findings reveal that frontier models remain vulnerable to agentic multimodal attacks, exposing a critical gap in current safety alignment. Warning: This paper contains potentially harmful content.

Autonomous coding agents increasingly execute multi-step software work, but lifecycle states such as reviewed, tested, DONE, and ready-to-merge remain claims unless supported by current evidence. We present Proof-or-Stop Lifecycle Control, a method that permits lifecycle transitions only when fresh, tracked-source-state-bound, mechanically verifiable evidence satisfies the relevant gate. The method treats agent outputs as claims rather than lifecycle state, and uses proof operationally to mean gate-admissible evidence under a stated trust model, not semantic program correctness. We evaluate an open-source implementation through mechanism tests, a powered control-policy ablation, and operated self-application evidence. The unattended-loop engine passed 10 of 10 scenarios with zero false-DONE, and local-key receipt bundles rejected 18 tamper classes with zero false accepts. In a 9,240-cell ablation, the pre-registered A4 versus A2-prime comparison reduced visible-pass/hidden-fail amplification from 31 of 1,800 injected cells under a compute-budgeted naive loop to 2 of 1,800 under the gated loop, a 1.6 percentage-point improvement in not-amplified rate with a 95 percent confidence interval of [0.8, 2.5]. A near-compute A3 versus A4 comparison, 14 of 1,800 versus 2 of 1,800, indicates that the gain is associated with enforcing review as a lifecycle gate rather than merely adding a reviewer. The self-application corpus contains 565 stories and 1,007 review findings, with 94.8 percent resolved, plus a 68-row high/critical cross-vendor exhibit. These results support Proof-or-Stop as a model-agnostic, host-neutral control layer for deciding which autonomous-agent claims a lifecycle may act on. The evaluation is limited to one model family, 24 ablation tasks, and a self-hosted corpus.

Background: Emerging reports of "AI psychosis" are on the rise, where user-LLM interactions may exacerbate or induce psychosis or adverse psychological symptoms. Whilst the sycophantic and agreeable nature of LLMs can be beneficial, it becomes a vector for harm by reinforcing delusional beliefs in vulnerable users. Methods: Psychosis-bench is a novel benchmark designed to systematically evaluate the psychogenicity of LLMs comprises 16 structured, 12-turn conversational scenarios simulating the progression of delusional themes(Erotic Delusions, Grandiose/Messianic Delusions, Referential Delusions) and potential harms. We evaluated eight prominent LLMs for Delusion Confirmation (DCS), Harm Enablement (HES), and Safety Intervention(SIS) across explicit and implicit conversational contexts. Findings: Across 1,536 simulated conversation turns, all LLMs demonstrated psychogenic potential, showing a strong tendency to perpetuate rather than challenge delusions (mean DCS of 0.91 pm0.88). Models frequently enabled harmful user requests (mean HES of 0.69 pm0.84) and offered safety interventions in only roughly a third of applicable turns (mean SIS of 0.37 pm0.48). 51 / 128 (39.8%) of scenarios had no safety interventions offered. Performance was significantly worse in implicit scenarios, models were more likely to confirm delusions and enable harm while offering fewer interventions (p < .001). A strong correlation was found between DCS and HES (rs = .77). Model performance varied widely, indicating that safety is not an emergent property of scale alone. Conclusion: This study establishes LLM psychogenicity as a quantifiable risk and underscores the urgent need for re-thinking how we train LLMs. We frame this issue not merely as a technical challenge but as a public health imperative requiring collaboration between developers, policymakers, and healthcare professionals.

Tool-using LLM agents increasingly coordinate real workloads by selecting and chaining third-party tools based on text-visible metadata such as tool names, descriptions, and return messages. We show that this convenience creates a supply-chain attack surface: a malicious MCP tool server can be co-registered alongside normal tools and induce overthinking loops, where individually trivial or plausible tool calls compose into cyclic trajectories that inflate end-to-end tokens and latency without any single step looking abnormal. We formalize this as a structural overthinking attack, distinguishable from token-level verbosity, and implement 14 malicious tools across three servers that trigger repetition, forced refinement, and distraction. Across heterogeneous registries and multiple tool-capable models, the attack causes severe resource amplification (up to 142.4times tokens) and can degrade task outcomes. Finally, we find that decoding-time concision controls do not reliably prevent loop induction, suggesting defenses should reason about tool-call structure rather than tokens alone.

Multimodal large language models with Retrieval Augmented Generation (RAG) have significantly advanced tasks such as multimodal question answering by grounding responses in external text and images. This grounding improves factuality, reduces hallucination, and extends reasoning beyond parametric knowledge. However, this reliance on external knowledge poses a critical yet underexplored safety risk: knowledge poisoning attacks, where adversaries deliberately inject adversarial multimodal content into external knowledge bases to steer model toward generating incorrect or even harmful responses. To expose such vulnerabilities, we propose MM-PoisonRAG, the first framework to systematically design knowledge poisoning in multimodal RAG. We introduce two complementary attack strategies: Localized Poisoning Attack (LPA), which implants targeted multimodal misinformation to manipulate specific queries, and Globalized Poisoning Attack (GPA), which inserts a single adversarial knowledge to broadly disrupt reasoning and induce nonsensical responses across all queries. Comprehensive experiments across tasks, models, and access settings show that LPA achieves targeted manipulation with attack success rates of up to 56%, while GPA completely disrupts model generation to 0% accuracy with just a single adversarial knowledge injection. Our results reveal the fragility of multimodal RAG and highlight the urgent need for defenses against knowledge poisoning.

Large language models (LLMs) have become integral to various real-world applications, leveraging massive, web-sourced datasets like Common Crawl, C4, and FineWeb for pretraining. While these datasets provide linguistic data essential for high-quality natural language generation, they often contain harmful content, such as hate speech, misinformation, and biased narratives. Training LLMs on such unfiltered data risks perpetuating toxic behaviors, spreading misinformation, and amplifying societal biases which can undermine trust in LLM-driven applications and raise ethical concerns about their use. This paper presents a large-scale analysis of inappropriate content across these datasets, offering a comprehensive taxonomy that categorizes harmful webpages into Topical and Toxic based on their intent. We also introduce a prompt evaluation dataset, a high-accuracy Topical and Toxic Prompt (TTP), and a transformer-based model (HarmFormer) for content filtering. Additionally, we create a new multi-harm open-ended toxicity benchmark (HAVOC) and provide crucial insights into how models respond to adversarial toxic inputs. Upon publishing, we will also opensource our model signal on the entire C4 dataset. Our work offers insights into ensuring safer LLM pretraining and serves as a resource for Responsible AI (RAI) compliance.

The multi-step, iterative image editing capabilities of multi-modal agentic systems have transformed digital content creation. Although latest image editing models faithfully follow instructions and generate high-quality images in single-turn edits, we identify a critical weakness in multi-turn editing, which is the iterative degradation of image quality. As images are repeatedly edited, minor artifacts accumulate, rapidly leading to a severe accumulation of visible noise and a failure to follow simple editing instructions. To systematically study these failures, we introduce Banana100, a comprehensive dataset of 28,000 degraded images generated through 100 iterative editing steps, including diverse textures and image content. Alarmingly, image quality evaluators fail to detect the degradation. Among 21 popular no-reference image quality assessment (NR-IQA) metrics, none of them consistently assign lower scores to heavily degraded images than to clean ones. The dual failures of generators and evaluators may threaten the stability of future model training and the safety of deployed agentic systems, if the low-quality synthetic data generated by multi-turn edits escape quality filters. We release the full code and data to facilitate the development of more robust models, helping to mitigate the fragility of multi-modal agentic systems.

Large Language Model (LLM) agents increasingly solve long-horizon tasks through multi-turn interactions with users and external tools. In these settings, relevant task information often unfolds over time rather than being fully specified at the initial prompt. Service agents make this challenge especially concrete: users may clarify or revise their goals, while tool responses provide information needed for subsequent decisions. Thus, a final reward alone cannot indicate which actions contributed to resolving the task. Recent methods rely on comparative evidence from other trajectories or resampled continuations, or on separately constructed step-level learning signals, to refine credit. However, a completed rollout already records how information and errors flow between agent actions. We introduce Influence-Aware Policy Optimization (IAPO), which represents each rollout as a typed influence-dependency graph over trainable agent actions, with user and tool observations serving as evidence. IAPO converts support-use and failed-use structure into routing weights that redistribute the same trajectory-level advantage. Experiments with Qwen3-4B and Qwen3-8B demonstrate superior performance over multi-turn reinforcement learning (RL) baselines across three service-agent benchmarks: {τ^2}-Bench, UserBench, and AgentChangeBench. BFCL-v4 Multi-Turn further shows that these gains do not compromise multi-turn function-calling performance. This work advances the understanding of credit assignment in multi-turn user interactions and provides a principled approach to training service agents from sparse outcome feedback.

Data poisoning attacks manipulate training data to introduce unexpected behaviors into machine learning models at training time. For text-to-image generative models with massive training datasets, current understanding of poisoning attacks suggests that a successful attack would require injecting millions of poison samples into their training pipeline. In this paper, we show that poisoning attacks can be successful on generative models. We observe that training data per concept can be quite limited in these models, making them vulnerable to prompt-specific poisoning attacks, which target a model's ability to respond to individual prompts. We introduce Nightshade, an optimized prompt-specific poisoning attack where poison samples look visually identical to benign images with matching text prompts. Nightshade poison samples are also optimized for potency and can corrupt an Stable Diffusion SDXL prompt in <100 poison samples. Nightshade poison effects "bleed through" to related concepts, and multiple attacks can composed together in a single prompt. Surprisingly, we show that a moderate number of Nightshade attacks can destabilize general features in a text-to-image generative model, effectively disabling its ability to generate meaningful images. Finally, we propose the use of Nightshade and similar tools as a last defense for content creators against web scrapers that ignore opt-out/do-not-crawl directives, and discuss possible implications for model trainers and content creators.

While Multi-Agent Systems (MAS) are increasingly deployed for complex workflows, their emergent properties-particularly the accumulation of bias-remain poorly understood. Because real-world MAS are too complex to analyze entirely, evaluating their ethical robustness requires first isolating their foundational mechanics. In this work, we conduct a baseline empirical study investigating how basic MAS topologies and feedback loops influence prejudice. Contrary to the assumption that multi-agent collaboration naturally dilutes bias, we hypothesize that structured workflows act as echo chambers, amplifying minor stochastic biases into systemic polarization. To evaluate this, we introduce Discrim-Eval-Open, an open-ended benchmark that bypasses individual model neutrality through forced comparative judgments across demographic groups. Analyzing bias cascades across various structures reveals that architectural sophistication frequently exacerbates bias rather than mitigating it. We observe systemic amplification even when isolated agents operate neutrally, and identify a 'Trigger Vulnerability' where injecting purely objective context drastically accelerates polarization. By stripping away advanced swarm complexity to study foundational dynamics, we establish a crucial baseline: structural complexity does not guarantee ethical robustness. Our code is available at https://github.com/weizhihao1/MAS-Bias.

Almost every paper on LLM jailbreaks and prompt injection reports an attack-success rate (ASR), and that number is assigned not by people but by an automated judge: either a safety classifier trained for the task, or a general chat model prompted to grade. The judge is rarely checked. We check it. Using 596 human-labeled completions from the HarmBench classifier validation set, we compare the two judge families against human majority votes and then attack them. The two families fail in opposite ways. The dedicated classifier over-flags (precision 0.835, recall 0.974); three different LLM-as-judges keep high precision (0.81 to 0.94) but show erratic recall (0.06 to 0.65), so the same responses produce very different ASR depending on which judge scores them. The two families also differ sharply in robustness. Wrappers that leave the harmful text untouched and only add benign framing flip every LLM-judge between 57% and 100% of the time, and a single prepended refusal sentence accounts for much of this (39% to 88%). The dedicated classifier resists these surface attacks (at most 6.7%), but a white-box GCG attack on its open weights flips 70% of confident true positives (21 of 30; 95% CI 54 to 86%) even at a small optimization budget. A two-annotator audit confirms the attacks leave the harm intact: every one of 80 sampled flips still contained the harmful content. Because a large and growing share of reported ASR comes from LLM-judges, many such numbers are unreliable both on average and under deliberate pressure. We recommend that papers report judge precision and recall on a human-labeled slice, report ASR corrected for judge precision, and include an adversarial check of the judge. Our code is released.

Large language models (LLMs) now reach expert-level scores on medical licensing exams, encouraging the assumption that high scores imply safe medical judgment while patients increasingly use them for health advice. We show this assumption is fragile: when misleading context is injected into questions that LLMs originally answer correctly, they abandon the correct answer. We call the ability to maintain correct judgment under adversarial context epistemic resilience, and introduce MedMisBench to measure it. MedMisBench contains 10,932 medical question items and 48,889 misleading context-option pairs spanning medical reasoning, agentic capability, and patient-journey evaluation. Across 11 model configurations, mean accuracy falls from 71.1% on original questions to 38.0% under focused misleading context, with 51.5% attack success. The most damaging injections are formal, rule-like fabrications: authority-framed falsehoods reach 69.5% attack success and exception-poisoning claims reach 64.1%. A 14-member clinical panel from 7 countries identified serious potential harm in 38.2% of reviewed cases. MedMisBench exposes a structural blind spot in LLM evaluation in medical settings: existing benchmarks measure what models know, but not whether they preserve correct medical judgment under misleading context.

Latent communication enables multi-agent systems to exchange information directly in internal representation space, reducing the token, computation, and latency overhead of text-based communication. To this end, lightweight trainable links are introduced to map the sender's representations into the receiver's input space. In this work, we show that even benign link training can increase harmful compliance relative to text-based communication while the underlying safety-aligned agents remain unchanged. An attacker can amplify this effect by optimizing the links on harmful query--response pairs or poisoning otherwise benign training data. We further develop a reinforcement-learning attack that rewards harmful compliance alongside benign task performance without requiring harmful target responses. Across three communication topologies and four safety benchmarks, this attack raises the mean harmful-compliance score from 27.9 with benignly trained links to 76.9. Compared with direct supervised optimization, it also achieves higher average accuracy on two benign utility benchmarks. Adapting the rewards toward safer behavior also enables repair of compromised links, substantially reducing harmful compliance across all evaluated attacks without updating the agents. Overall, our results show that safety alignment requires considering the multi-agent system as a whole.

While state-of-the-art diffusion models (DMs) excel in image generation, concerns regarding their security persist. Earlier research highlighted DMs' vulnerability to data poisoning attacks, but these studies placed stricter requirements than conventional methods like `BadNets' in image classification. This is because the art necessitates modifications to the diffusion training and sampling procedures. Unlike the prior work, we investigate whether BadNets-like data poisoning methods can directly degrade the generation by DMs. In other words, if only the training dataset is contaminated (without manipulating the diffusion process), how will this affect the performance of learned DMs? In this setting, we uncover bilateral data poisoning effects that not only serve an adversarial purpose (compromising the functionality of DMs) but also offer a defensive advantage (which can be leveraged for defense in classification tasks against poisoning attacks). We show that a BadNets-like data poisoning attack remains effective in DMs for producing incorrect images (misaligned with the intended text conditions). Meanwhile, poisoned DMs exhibit an increased ratio of triggers, a phenomenon we refer to as `trigger amplification', among the generated images. This insight can be then used to enhance the detection of poisoned training data. In addition, even under a low poisoning ratio, studying the poisoning effects of DMs is also valuable for designing robust image classifiers against such attacks. Last but not least, we establish a meaningful linkage between data poisoning and the phenomenon of data replications by exploring DMs' inherent data memorization tendencies.

Recent research has highlighted the vulnerability of Deep Neural Networks (DNNs) against data poisoning attacks. These attacks aim to inject poisoning samples into the models' training dataset such that the trained models have inference failures. While previous studies have executed different types of attacks, one major challenge that greatly limits their effectiveness is the uncertainty of the re-training process after the injection of poisoning samples, including the re-training initialization or algorithms. To address this challenge, we propose a novel attack method called ''Sharpness-Aware Data Poisoning Attack (SAPA)''. In particular, it leverages the concept of DNNs' loss landscape sharpness to optimize the poisoning effect on the worst re-trained model. It helps enhance the preservation of the poisoning effect, regardless of the specific retraining procedure employed. Extensive experiments demonstrate that SAPA offers a general and principled strategy that significantly enhances various types of poisoning attacks.

Current evaluations of LLM safety predominantly rely on severity-based taxonomies to assess the harmfulness of malicious queries. We argue that this formulation requires re-examination as it assumes uniform risk across all malicious queries, neglecting Execution Likelihood--the conditional probability of a threat being realized given the model's response. In this work, we introduce Expected Harm, a metric that weights the severity of a jailbreak by its execution likelihood, modeled as a function of execution cost. Through empirical analysis of state-of-the-art models, we reveal a systematic Inverse Risk Calibration: models disproportionately exhibit stronger refusal behaviors for low-likelihood (high-cost) threats while remaining vulnerable to high-likelihood (low-cost) queries. We demonstrate that this miscalibration creates a structural vulnerability: by exploiting this property, we increase the attack success rate of existing jailbreaks by up to 2times. Finally, we trace the root cause of this failure using linear probing, which reveals that while models encode severity in their latent space to drive refusal decisions, they possess no distinguishable internal representation of execution cost, making them "blind" to this critical dimension of risk.

The rise of Large Audio Language Models (LAMs) brings both potential and risks, as their audio outputs may contain harmful or unethical content. However, current research lacks a systematic, quantitative evaluation of LAM safety especially against jailbreak attacks, which are challenging due to the temporal and semantic nature of speech. To bridge this gap, we introduce AJailBench, the first benchmark specifically designed to evaluate jailbreak vulnerabilities in LAMs. We begin by constructing AJailBench-Base, a dataset of 1,495 adversarial audio prompts spanning 10 policy-violating categories, converted from textual jailbreak attacks using realistic text to speech synthesis. Using this dataset, we evaluate several state-of-the-art LAMs and reveal that none exhibit consistent robustness across attacks. To further strengthen jailbreak testing and simulate more realistic attack conditions, we propose a method to generate dynamic adversarial variants. Our Audio Perturbation Toolkit (APT) applies targeted distortions across time, frequency, and amplitude domains. To preserve the original jailbreak intent, we enforce a semantic consistency constraint and employ Bayesian optimization to efficiently search for perturbations that are both subtle and highly effective. This results in AJailBench-APT, an extended dataset of optimized adversarial audio samples. Our findings demonstrate that even small, semantically preserved perturbations can significantly reduce the safety performance of leading LAMs, underscoring the need for more robust and semantically aware defense mechanisms.

Large Language Models (LLMs) are typically evaluated for safety under single-shot or low-budget adversarial prompting, which underestimates real-world risk. In practice, attackers can exploit large-scale parallel sampling to repeatedly probe a model until a harmful response is produced. While recent work shows that attack success increases with repeated sampling, principled methods for predicting large-scale adversarial risk remain limited. We propose a scaling-aware Best-of-N estimation of risk, SABER, for modeling jailbreak vulnerability under Best-of-N sampling. We model sample-level success probabilities using a Beta distribution, the conjugate prior of the Bernoulli distribution, and derive an analytic scaling law that enables reliable extrapolation of large-N attack success rates from small-budget measurements. Using only n=100 samples, our anchored estimator predicts ASR@1000 with a mean absolute error of 1.66, compared to 12.04 for the baseline, which is an 86.2% reduction in estimation error. Our results reveal heterogeneous risk scaling profiles and show that models appearing robust under standard evaluation can experience rapid nonlinear risk amplification under parallel adversarial pressure. This work provides a low-cost, scalable methodology for realistic LLM safety assessment. We will release our code and evaluation scripts upon publication to future research.

Adapting reasoning models to new tasks during post-training with only output-level supervision stalls under reinforcement learning from verifiable rewards (RLVR) when the initial success probability p_0 is small. Using the Tsallis q-logarithm, we define a loss family J_Q that interpolates between RLVR (at q{=}0, the exploitation pole) and the log-marginal-likelihood over latent trajectories (at q{=}1, the density-estimation pole). All members share the same per-example gradient direction, differing only by a scalar amplification P_{θ^{-q}} that reweights each instance independently of the learning rate. This amplification is the mechanism that addresses cold-start stalling: under gradient flow, the exploitation pole requires Ω(1{p_0}) time to escape cold start, while the density-estimation pole escapes in Θbig(log(1{p_0})big); intermediate q trades escape speed against noise memorization. Because P_θ is intractable, we derive two Monte Carlo estimators from the two factorizations of the gradient: Gradient-Amplified RL (GARL) samples from the prior and amplifies the RL gradient, and Posterior-Attenuated Fine-Tuning (PAFT) importance-resamples from the posterior and runs standard SFT. Both have bias Obig(q{M P_θ^{q+1}}big); GARL has lower variance, PAFT has semantically coherent gradients. On FinQA, HotPotQA, and MuSiQue, GARL at q{=}0.75 substantially mitigates cold-start stalling, escaping cold start where GRPO fails entirely. In warm start, GARL at low q dominates FinQA where training is stable; on HotPotQA and MuSiQue, GARL destabilizes during training, and PAFT at q{=}0.75 provides stable gradients (best overall on HotPotQA at 47.9 maj@16, +14.4 over GRPO).

LLM rerankers add of the order of \$0.2-0.3 per 1,000 queries and about a second of tail latency on top of a graph-augmented dense pipeline such as HippoRAG2, and on three multi-hop benchmarks they improve final-hop top-K coverage on seven of nine (dataset, K) cells, by up to +34.8 pp. We ask whether a learned per-query gate can skip the reranker where it will not help, using only features available before the LLM call (27 score and lexical statistics of the two retrieval lists plus a PCA of a small query embedding) with an executable fallback. Every choice, including the fallback and the threshold, is made inside the training fold and applied once to held-out queries, and harmful skips (the rerank would have found the target, the fallback did not) are reported next to the aggregate coverage. Across nine cells on 2WikiMultiHopQA, MuSiQue and HotpotQA the gate skips 51% of calls at an average held-out LastHop@K cost of 1.2 pp; four cells meet a pre-registered 1 pp rule, harmful skips occur in eight (190 harmful against 136 beneficial), and a random gate at the same skip rate loses 2 to 11 pp on the high-lift cells. A second rule sets each cell's threshold from a pre-specified budget on the expected harmful-skip rate over Platt-calibrated harm probabilities (ECE 0.025 after calibration, 0.094 before): at a 1 pp budget the gate skips 42% at -0.8 pp with 66 harmful skips and six cells within 1 pp, but realised harm exceeds the promise in six cells (mean 1.45 vs 0.83 pp), a selection optimism we quantify; a 0.5 pp budget realises about 1 pp. The harm probabilities are calibrated but barely discriminative (AUC 0.16 to 0.70). An earlier version reported 73% "lossless" savings; that figure rested on an oracle fallback and a wrong MuSiQue target, and we document both.

Reinforcement Learning with Verifiable Rewards (RLVR) is widely used to improve reasoning in multiple domains, yet outcome-only scalar rewards are often sparse and uninformative, especially on failed samples, where they merely indicate failure and provide no insight into why the reasoning fails. In this paper, we investigate how to leverage richer verbal feedback to guide RLVR training on failed samples, and how to convert such feedback into a trainable learning signal. Specifically, we propose a multi-turn feedback-guided reinforcement learning framework. It builds on three mechanisms: (1) dynamic multi-turn regeneration guided by feedback, triggered only on failed samples, (2) two complementary learning signals for within-turn and cross-turn optimization, and (3) structured feedback injection into the model's reasoning process. Trained on sampled OpenR1-Math, the approach outperforms supervised fine-tuning and RLVR baselines in-domain and generalizes well out-of-domain.

Safety benchmarks such as HarmBench rely on LLM judges to classify model responses as harmful or safe, yet the judge configuration, namely the combination of judge model and judge prompt, is typically treated as a fixed implementation detail. We show this assumption is problematic. Using a 2 x 2 x 3 factorial design, we construct 12 judge prompt variants along two axes, evaluation structure and instruction framing, and apply them using a single judge model, Claude Sonnet 4-6, producing 28,812 judgments over six target models and 400 HarmBench behaviors. We find that prompt wording alone, holding the judge model fixed, shifts measured harmful-response rates by up to 24.2 percentage points, with even within-condition surface rewording causing swings of up to 20.1 percentage points. Model safety rankings are moderately unstable, with mean Kendall tau = 0.89, and category-level sensitivity ranges from 39.6 percentage points for copyright to 0 percentage points for harassment. A supplementary multi-judge experiment using three judge models shows that judge-model choice adds further variance. Our results demonstrate that judge prompt wording is a substantial, previously under-examined source of measurement variance in safety benchmarking.

Jailbreak attacks pose significant threats to large language models (LLMs), enabling attackers to bypass safeguards. However, existing reactive defense approaches struggle to keep up with the rapidly evolving multi-turn jailbreaks, where attackers continuously deepen their attacks to exploit vulnerabilities. To address this critical challenge, we propose HoneyTrap, a novel deceptive LLM defense framework leveraging collaborative defenders to counter jailbreak attacks. It integrates four defensive agents, Threat Interceptor, Misdirection Controller, Forensic Tracker, and System Harmonizer, each performing a specialized security role and collaborating to complete a deceptive defense. To ensure a comprehensive evaluation, we introduce MTJ-Pro, a challenging multi-turn progressive jailbreak dataset that combines seven advanced jailbreak strategies designed to gradually deepen attack strategies across multi-turn attacks. Besides, we present two novel metrics: Mislead Success Rate (MSR) and Attack Resource Consumption (ARC), which provide more nuanced assessments of deceptive defense beyond conventional measures. Experimental results on GPT-4, GPT-3.5-turbo, Gemini-1.5-pro, and LLaMa-3.1 demonstrate that HoneyTrap achieves an average reduction of 68.77% in attack success rates compared to state-of-the-art baselines. Notably, even in a dedicated adaptive attacker setting with intensified conditions, HoneyTrap remains resilient, leveraging deceptive engagement to prolong interactions, significantly increasing the time and computational costs required for successful exploitation. Unlike simple rejection, HoneyTrap strategically wastes attacker resources without impacting benign queries, improving MSR and ARC by 118.11% and 149.16%, respectively.

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.

Tool-use capability is a fundamental component of LLM agents, enabling them to interact with external systems through structured function calls. However, existing research exhibits inconsistent interaction representations, largely overlooks the structural distribution of tool-use trajectories, and relies on incompatible evaluation benchmarks. We present UniToolCall, a unified framework for tool learning that standardizes the entire pipeline from toolset construction and dataset generation to evaluation. The framework curates a large tool pool of 22k+ tools and constructs a hybrid training corpus of 390k+ instances by combining 10 standardized public datasets with structurally controlled synthetic trajectories. It explicitly models diverse interaction patterns, including single-hop vs. multi-hop and single-turn vs. multi-turn, while capturing both serial and parallel execution structures. To support coherent multi-turn reasoning, we further introduce an Anchor Linkage mechanism that enforces cross-turn dependencies. Furthermore, we convert 7 public benchmarks into a unified Query--Action--Observation--Answer (QAOA) representation with fine-grained evaluation at the function-call, turn, and conversation levels. Experiments show that fine-tuning Qwen3-8B on our dataset substantially improves tool-use performance. Under the distractor-heavy Hybrid-20 setting, achieves 93.0% single-turn Strict Precision, outperforming commercial models including GPT, Gemini, and Claude.

Large Language Models (LLMs) with safe-alignment training are powerful instruments with robust language comprehension capabilities. These models typically undergo meticulous alignment procedures involving human feedback to ensure the acceptance of safe inputs while rejecting harmful or unsafe ones. However, despite their massive scale and alignment efforts, LLMs remain vulnerable to jailbreak attacks, where malicious users manipulate the model to produce harmful outputs that it was explicitly trained to avoid. In this study, we find that the safety mechanisms in LLMs are predominantly embedded in the middle-to-late layers. Building on this insight, we introduce a novel white-box jailbreak method, SABER (Safety Alignment Bypass via Extra Residuals), which connects two intermediate layers s and e such that s < e, through a residual connection. Our approach achieves a 51% improvement over the best-performing baseline on the HarmBench test set. Furthermore, SABER induces only a marginal shift in perplexity when evaluated on the HarmBench validation set. The source code is publicly available at https://github.com/PalGitts/SABER.

Multi-attribution learning (MAL), which enhances model performance by learning from conversion labels yielded by multiple attribution mechanisms, has emerged as a promising learning paradigm for conversion rate (CVR) prediction. However, the conversion labels in public CVR datasets are generated by a single attribution mechanism, hindering the development of MAL approaches. To address this data gap, we establish the Multi-Attribution Benchmark (MAC), the first public CVR dataset featuring labels from multiple attribution mechanisms. Besides, to promote reproducible research on MAL, we develop PyMAL, an open-source library covering a wide array of baseline methods. We conduct comprehensive experimental analyses on MAC and reveal three key insights: (1) MAL brings consistent performance gains across different attribution settings, especially for users featuring long conversion paths. (2) The performance growth scales up with objective complexity in most settings; however, when predicting first-click conversion targets, simply adding auxiliary objectives is counterproductive, underscoring the necessity of careful selection of auxiliary objectives. (3) Two architectural design principles are paramount: first, to fully learn the multi-attribution knowledge, and second, to fully leverage this knowledge to serve the main task. Motivated by these findings, we propose Mixture of Asymmetric Experts (MoAE), an effective MAL approach incorporating multi-attribution knowledge learning and main task-centric knowledge utilization. Experiments on MAC show that MoAE substantially surpasses the existing state-of-the-art MAL method. We believe that our benchmark and insights will foster future research in the MAL field. Our MAC benchmark and the PyMAL algorithm library are publicly available at https://github.com/alimama-tech/PyMAL.

Existing machine learning methods for causal inference usually estimate quantities expressed via the mean of potential outcomes (e.g., average treatment effect). However, such quantities do not capture the full information about the distribution of potential outcomes. In this work, we estimate the density of potential outcomes after interventions from observational data. For this, we propose a novel, fully-parametric deep learning method called Interventional Normalizing Flows. Specifically, we combine two normalizing flows, namely (i) a nuisance flow for estimating nuisance parameters and (ii) a target flow for parametric estimation of the density of potential outcomes. We further develop a tractable optimization objective based on a one-step bias correction for efficient and doubly robust estimation of the target flow parameters. As a result, our Interventional Normalizing Flows offer a properly normalized density estimator. Across various experiments, we demonstrate that our Interventional Normalizing Flows are expressive and highly effective, and scale well with both sample size and high-dimensional confounding. To the best of our knowledge, our Interventional Normalizing Flows are the first proper fully-parametric, deep learning method for density estimation of potential outcomes.

Large language models (LLMs) are increasingly deployed in real-world applications, raising concerns about their security. While jailbreak attacks highlight failures under overtly harmful queries, they overlook a critical risk: incorrectly answering harmless-looking inputs can be dangerous and cause real-world harm (Implicit Harm). We systematically reformulate the LLM risk landscape through a structured quadrant perspective based on output factuality and input harmlessness, uncovering an overlooked high-risk region. To investigate this gap, we propose JailFlipBench, a benchmark aims to capture implicit harm, spanning single-modal, multimodal, and factual extension scenarios with diverse evaluation metrics. We further develop initial JailFlip attack methodologies and conduct comprehensive evaluations across multiple open-source and black-box LLMs, show that implicit harm present immediate and urgent real-world risks, calling for broader LLM safety assessments and alignment beyond conventional jailbreak paradigms.

Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) systems are increasingly integrated into clinical workflows; however, prompt injection attacks can steer these systems toward clinically unsafe or misleading outputs. We introduce the Medical Prompt Injection Benchmark (MPIB), a dataset-and-benchmark suite for evaluating clinical safety under both direct prompt injection and indirect, RAG-mediated injection across clinically grounded tasks. MPIB emphasizes outcome-level risk via the Clinical Harm Event Rate (CHER), which measures high-severity clinical harm events under a clinically grounded taxonomy, and reports CHER alongside Attack Success Rate (ASR) to disentangle instruction compliance from downstream patient risk. The benchmark comprises 9,697 curated instances constructed through multi-stage quality gates and clinical safety linting. Evaluating MPIB across a diverse set of baseline LLMs and defense configurations, we find that ASR and CHER can diverge substantially, and that robustness depends critically on whether adversarial instructions appear in the user query or in retrieved context. We release MPIB with evaluation code, adversarial baselines, and comprehensive documentation to support reproducible and systematic research on clinical prompt injection. Code and data are available at GitHub (code) and Hugging Face (data).

Recent advances in end-to-end trained omni-models have substantially improved audio capabilities by strengthening text-audio modality alignment. However, whether such alignment inadvertently facilitates the transfer of safety vulnerabilities across modalities remains underexplored. This question is critical as text-based jailbreak attacks are considerably more mature than audio-based ones; if they transfer systematically, current audio safety evaluations may underestimate risks originating from the text modality. In this paper, we introduce the Alignment Curse, a formally characterized and empirically validated principle showing that stronger modality alignment enables more effective transfer of attacks from text to audio, revealing a fundamental tension between capability and safety. Motivated by this principle, we conduct a comprehensive black-box evaluation of three attack categories on recent omni-models (e.g., Qwen2.5-Omni, Qwen3-Omni): text attacks, text-transferred audio attacks, and audio attacks. We find that text-transferred audio attacks perform comparably to, and often better than, audio-based attacks, exhibiting a clear advantage under audio-only access. This suggests that text-based vulnerabilities play a pivotal role in shaping audio safety risks. Finally, we empirically analyze the relationship between modality alignment and transfer effectiveness across attack methods and models, observing consistent support for the Alignment Curse: tighter modality alignment leads to more effective cross-modality attack transfer.

Multimodal Large Language Models (MLLMs) are increasingly used for content moderation, yet their robustness in short-form video contexts remains underexplored. Current safety evaluations often rely on unimodal attacks, failing to address combined attack vulnerabilities. In this paper, we introduce a comprehensive framework for evaluating the tri-modal safety of MLLMs. First, we present the Short-Video Multimodal Adversarial (SVMA) dataset, comprising diverse short-form videos with human-guided synthetic adversarial attacks. Second, we propose ChimeraBreak, a novel tri-modal attack strategy that simultaneously challenges visual, auditory, and semantic reasoning pathways. Extensive experiments on state-of-the-art MLLMs reveal significant vulnerabilities with high Attack Success Rates (ASR). Our findings uncover distinct failure modes, showing model biases toward misclassifying benign or policy-violating content. We assess results using LLM-as-a-judge, demonstrating attack reasoning efficacy. Our dataset and findings provide crucial insights for developing more robust and safe MLLMs.

Organizations across finance, healthcare, transportation, content moderation, and critical infrastructure are rapidly deploying highly automated AI systems, yet they lack principled methods to quantify how increasing automation amplifies harm when failures occur. We propose a parsimonious Bayesian risk decomposition expressing expected loss as the product of three terms: the probability of system failure, the conditional probability that a failure propagates into harm given the automation level, and the expected severity of harm. This framework isolates a critical quantity -- the conditional probability that failures propagate into harm -- which captures execution and oversight risk rather than model accuracy alone. We develop complete theoretical foundations: formal proofs of the decomposition, a harm propagation equivalence theorem linking the harm propagation probability to observable execution controls, risk elasticity measures, efficient frontier analysis for automation policy, and optimal resource allocation principles with second-order conditions. We motivate the framework with an illustrative case study of the 2012 Knight Capital incident ($440M loss) as one instantiation of a broadly applicable failure pattern, and characterize the research design required to empirically validate the framework at scale across deployment domains. This work provides the theoretical foundations for a new class of deployment-focused risk governance tools for agentic and automated AI systems.

As computer vision systems become more widely deployed, there is increasing concern from both the research community and the public that these systems are not only reproducing but amplifying harmful social biases. The phenomenon of bias amplification, which is the focus of this work, refers to models amplifying inherent training set biases at test time. Existing metrics measure bias amplification with respect to single annotated attributes (e.g., computer). However, several visual datasets consist of images with multiple attribute annotations. We show models can learn to exploit correlations with respect to multiple attributes (e.g., {computer, keyboard}), which are not accounted for by current metrics. In addition, we show current metrics can give the erroneous impression that minimal or no bias amplification has occurred as they involve aggregating over positive and negative values. Further, these metrics lack a clear desired value, making them difficult to interpret. To address these shortcomings, we propose a new metric: Multi-Attribute Bias Amplification. We validate our proposed metric through an analysis of gender bias amplification on the COCO and imSitu datasets. Finally, we benchmark bias mitigation methods using our proposed metric, suggesting possible avenues for future bias mitigation

Large language models (LLMs) are improving at an exceptional rate. However, these models are still susceptible to jailbreak attacks, which are becoming increasingly dangerous as models become increasingly powerful. In this work, we introduce a dataset of jailbreaks where each example can be input in both a single or a multi-turn format. We show that while equivalent in content, they are not equivalent in jailbreak success: defending against one structure does not guarantee defense against the other. Similarly, LLM-based filter guardrails also perform differently depending on not just the input content but the input structure. Thus, vulnerabilities of frontier models should be studied in both single and multi-turn settings; this dataset provides a tool to do so.

As large language models (LLMs) become high-privilege agents in risk-sensitive settings, they introduce systemic threats beyond hallucination, where minor compliance errors can cause critical data leaks. However, existing benchmarks focus on rule-based QA, lacking agentic execution modeling, overlooking compliance drift in adversarial interactions, and relying on binary safety metrics that fail to capture behavioral degradation. To bridge these gaps, we present CNFinBench, a comprehensive benchmark spanning 29 subtasks grounded in the triad of expertise, autonomy, and integrity. It assesses domain-specific capabilities through certified regulatory corpora and professional financial tasks, reconstructs end-to-end agent workflows from requirement parsing to tool verification, and simulates multi-turn adversarial attacks that induce behavioral compliance drift. To quantify safety degradation, we introduce the Harmful Instruction Compliance Score (HICS), a multi-dimensional safety metric that integrates risk-type-specific deductions, multi-turn consistency tracking, and severity-adjusted penalty scaling based on fine-grained violation triggers. Evaluations over 22 open-/closed-source models reveal: LLMs perform well in applied tasks yet lack robust rule understanding, suffer a 15.4-point drop single modules to full execution chains, and collapse rapidly in multi-turn attacks, with average violations surging by 172.3% in Round 2. CNFinBench is available at https://cnfinbench.opencompass.org.cn and https://github.com/VertiAIBench/CNFinBench.

Algorithmic fairness has attracted significant attention in recent years, with many quantitative measures suggested for characterizing the fairness of different machine learning algorithms. Despite this interest, the robustness of those fairness measures with respect to an intentional adversarial attack has not been properly addressed. Indeed, most adversarial machine learning has focused on the impact of malicious attacks on the accuracy of the system, without any regard to the system's fairness. We propose new types of data poisoning attacks where an adversary intentionally targets the fairness of a system. Specifically, we propose two families of attacks that target fairness measures. In the anchoring attack, we skew the decision boundary by placing poisoned points near specific target points to bias the outcome. In the influence attack on fairness, we aim to maximize the covariance between the sensitive attributes and the decision outcome and affect the fairness of the model. We conduct extensive experiments that indicate the effectiveness of our proposed attacks.

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.

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.

Drug-side effect research is vital for understanding adverse reactions arising in complex multi-drug therapies. However, the scarcity of higher-order datasets that capture the combinatorial effects of multiple drugs severely limits progress in this field. Existing resources such as TWOSIDES primarily focus on pairwise interactions. To fill this critical gap, we introduce HODDI, the first Higher-Order Drug-Drug Interaction Dataset, constructed from U.S. Food and Drug Administration (FDA) Adverse Event Reporting System (FAERS) records spanning the past decade, to advance computational pharmacovigilance. HODDI contains 109,744 records involving 2,506 unique drugs and 4,569 unique side effects, specifically curated to capture multi-drug interactions and their collective impact on adverse effects. Comprehensive statistical analyses demonstrate HODDI's extensive coverage and robust analytical metrics, making it a valuable resource for studying higher-order drug relationships. Evaluating HODDI with multiple models, we found that simple Multi-Layer Perceptron (MLP) can outperform graph models, while hypergraph models demonstrate superior performance in capturing complex multi-drug interactions, further validating HODDI's effectiveness. Our findings highlight the inherent value of higher-order information in drug-side effect prediction and position HODDI as a benchmark dataset for advancing research in pharmacovigilance, drug safety, and personalized medicine. The dataset and codes are available at https://github.com/TIML-Group/HODDI.

Speech-to-speech models handle turn-taking naturally but offer limited support for tool-calling or complex reasoning, while production ASR-LLM-TTS voice pipelines offer these capabilities but rely on silence timeouts, which lead to unnatural turn-taking. We present DualTurn, which narrows this gap through generative pretraining on dual-channel conversational audio. The model generates both speakers' future audio autoregressively, implicitly learning conversational dynamics without any labels, and is then fine-tuned to predict interpretable turn-taking signals that map directly to agent actions. DualTurn monitors both channels continuously, anticipating turn boundaries and producing five agent actions. On standard benchmarks, DualTurn (0.5B) outperforms both VAP on agent action prediction (wF1 0.633 vs. 0.389) and a 3.1B audio-text model on word-level turn prediction (AUC 0.930 vs. 0.880), while anticipating turn boundaries earlier with fewer interruptions.

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.

With the rapid advancement of multimodal large language models (MLLMs), concerns regarding their security have increasingly captured the attention of both academia and industry. Although MLLMs are vulnerable to jailbreak attacks, designing effective multimodal jailbreak attacks poses unique challenges, especially given the distinct protective measures implemented across various modalities in commercial models. Previous works concentrate risks into a single modality, resulting in limited jailbreak performance. In this paper, we propose a heuristic-induced multimodal risk distribution jailbreak attack method, called HIMRD, which consists of two elements: multimodal risk distribution strategy and heuristic-induced search strategy. The multimodal risk distribution strategy is used to segment harmful instructions across multiple modalities to effectively circumvent MLLMs' security protection. The heuristic-induced search strategy identifies two types of prompts: the understanding-enhancing prompt, which helps the MLLM reconstruct the malicious prompt, and the inducing prompt, which increases the likelihood of affirmative outputs over refusals, enabling a successful jailbreak attack. Extensive experiments demonstrate that this approach effectively uncovers vulnerabilities in MLLMs, achieving an average attack success rate of 90% across seven popular open-source MLLMs and an average attack success rate of around 68% in three popular closed-source MLLMs. Our code will coming soon. Warning: This paper contains offensive and harmful examples, reader discretion is advised.

In recent years, unified multimodal models (UMMs) have emerged to support both understanding and generation within a single framework. Mastering dynamic, multi-turn interleaved image-text dialogues is a crucial task for UMMs in real-world applications. However, existing benchmarks fail to evaluate this important task, as they are often limited to single-turn or static settings, and typically overlook exposure bias in multi-turn interactions. To bridge this gap, we propose IMUG-Bench, a comprehensive benchmark for multi-turn interleaved image-text dialogue of UMMs that jointly evaluates their understanding and generation capabilities. Our IMUG-Bench comprises three classes: Static Spatial, Temporal Causal, and Hybrid, covering 3,113 samples and 12,034 interaction turns. It also includes dynamic understanding questions, thereby supporting evaluation that better reflects real-world multi-turn interaction scenarios. Large-scale experiments on IMUG-Bench systematically evaluate mainstream open-source and closed-source UMMs, revealing their capability boundaries and failure modes, and uncovering pronounced exposure bias on the generation side in multi-turn interactions. We further explore several test-time scaling strategies, including Chain-of-Thought, Self-Verification, and Best-of-N Sampling, which effectively improve generation accuracy and mitigate exposure bias in generation tasks. These findings provide insights into enhancing the robustness and multi-turn interaction capability of future UMMs.

Activation steering has emerged as a powerful tool to shape LLM behavior without the need for weight updates. While its inherent brittleness and unreliability are well-documented, its safety implications remain underexplored. In this work, we present a systematic safety audit of steering vectors obtained with Contrastive Activation Addition (CAA), a widely used steering approach, under a unified evaluation protocol. Using JailbreakBench as benchmark, we show that steering vectors consistently influence the success rate of jailbreak attacks, with stronger amplification under simple template-based attacks. Across LLM families and sizes, steering the model in specific directions can drastically increase (up to 57%) or decrease (up to 50%) its attack success rate (ASR), depending on the targeted behavior. We attribute this phenomenon to the overlap between the steering vectors and the latent directions of refusal behavior. Thus, we offer a traceable explanation for this discovery. Together, our findings reveal the previously unobserved origin of this safety gap in LLMs, highlighting a trade-off between controllability and safety.

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.

Hardware faults can flip bits in the stored weights of a quantized neural network, potentially compromising its predictions. While such faults typically affect multiple parameters simultaneously, existing verifiers are limited to single-parameter perturbations due to the combinatorial explosion of possible flip locations in large networks. We present mBFV (m-BitFlip Verifier), an efficient verification framework that proves robustness against simultaneous bit flips across multiple parameters without explicitly enumerating these combinations. mBFV achieves this via a novel multi-parameter bound propagation technique that directly aggregates the m worst-case contributions. To further tighten these bounds, mBFV employs a branch-and-bound mechanism over perturbation locations, partitioning the potential flips to smaller groups of neurons. Evaluated on 625 instances, mBFV successfully verifies 293, significantly outperforming a prior single-parameter verifier (38 verified instances) and an exact mixed-integer linear programming baseline (0 verified instances). Notably, while these baselines are restricted to single-parameter flips on small networks (up to 13k parameters), mBFV scales to verify networks with up to 1.15M parameters against up to four simultaneous parameter flips.

The agent--tool interaction loop is a critical attack surface for modern Large Language Model (LLM) agents. Existing denial-of-service (DoS) attacks typically function at the user-prompt or retrieval-augmented generation (RAG) context layer and are inherently single-turn in nature. This limitation restricts cost amplification and diminishes stealth in goal-oriented workflows. To address these issues, we proposed a stealthy, multi-turn economic DoS attack at the tool layer under the Model Context Protocol (MCP). By simply editing text-visible fields and implementing a template-driven return policy, our malicious server preserves function signatures and the terminal benign payload while steering agents into prolonged, verbose tool-calling chains. We optimize these text-only edits with Monte Carlo Tree Search (MCTS) to maximize cost under a task-success constraint. Across six LLMs on ToolBench and BFCL benchmarks, our attack yields trajectories over 60K tokens, increases per-query cost by up to 658 times, raises energy by 100 to 560 times, and pushes GPU key-value (KV) cache occupancy to 35--74%. Standard prompt filters and output trajectory monitors seldom detect these attacks, highlighting the need for defenses that safeguard agentic processes rather than focusing solely on final outcomes. We will release the code soon.

Multimodal Large Language Models (MLLMs) have achieved impressive performance and have been put into practical use in commercial applications, but they still have potential safety mechanism vulnerabilities. Jailbreak attacks are red teaming methods that aim to bypass safety mechanisms and discover MLLMs' potential risks. Existing MLLMs' jailbreak methods often bypass the model's safety mechanism through complex optimization methods or carefully designed image and text prompts. Despite achieving some progress, they have a low attack success rate on commercial closed-source MLLMs. Unlike previous research, we empirically find that there exists a Shuffle Inconsistency between MLLMs' comprehension ability and safety ability for the shuffled harmful instruction. That is, from the perspective of comprehension ability, MLLMs can understand the shuffled harmful text-image instructions well. However, they can be easily bypassed by the shuffled harmful instructions from the perspective of safety ability, leading to harmful responses. Then we innovatively propose a text-image jailbreak attack named SI-Attack. Specifically, to fully utilize the Shuffle Inconsistency and overcome the shuffle randomness, we apply a query-based black-box optimization method to select the most harmful shuffled inputs based on the feedback of the toxic judge model. A series of experiments show that SI-Attack can improve the attack's performance on three benchmarks. In particular, SI-Attack can obviously improve the attack success rate for commercial MLLMs such as GPT-4o or Claude-3.5-Sonnet.

Mixture-of-Experts (MoE) architectures in Large Language Models (LLMs) have significantly reduced inference costs through sparse activation. However, this sparse activation paradigm also introduces new safety challenges. Since only a subset of experts is engaged for each input, model behavior becomes coupled to routing decisions, yielding a difficult-to-control mechanism that can vary across safety-relevant scenarios. At the same time, adapting model behavior through full fine-tuning or retraining is costly, especially when developers need to rapidly configure the same model for different safety objectives. We present MASCing (MoE Activation Steering Configuration), the first framework that enables flexible reconfiguration of MoE behavior across diverse safety scenarios without retraining. MASCing uses an LSTM-based surrogate model to capture cross-layer routing dependencies and map routing logits to downstream behaviors. It then optimizes a steering matrix to identify behavior-relevant expert circuits and, at inference time, applies steering masks to the routing gates to override expert selection. This enables targeted enhancement or suppression of specific behaviors while preserving general language utility. To demonstrate its reconfigurability, we apply MASCing to two different safety-related objectives and observe consistent gains with negligible overhead across seven open-source MoE models. For multi-turn jailbreak defense, it improves the average defense success rate from 52.5% to 83.9%, with gains of up to 89.2%. For adult-content generation, MASCing enables models to comply with such requests that would otherwise be refused, increasing the average generation success rate from 52.6% to 82.0%, with gains of up to 93.0%. These results establish MASCing as a practical, lightweight, and flexible framework for scenario-specific safety reconfiguration in MoE models.

Red-team evaluations of AI models support some claims and not others, and the boundary between the two is calculable rather than merely a matter of judgment. We define the evidential ceiling of an evaluation as the largest factor by which one result can move belief under a fixed testing budget, derive it in closed form for the benchmark null result, and use it to locate that boundary exactly. We find that above a calculable harm rate, a benchmark of modest size certifies a category to a stated evidentiary standard, and a clean sheet is then the stronger of the two possible observations, outweighing a single reproduced failure. Below that rate, no passive benchmark of feasible size provides the specified evidence of safety under the fixed scoring rule and approximately independent trial structure. The crossing between the two regimes has a closed form. The bound is not specific to benchmarks: written in terms of a procedure's hypothesis conditioned elicitation rates, it covers adaptive and automated red teaming as well, and shows that discrimination between the hypotheses rather than attack success is what determines evidential worth. Auditing eight evaluation suites against the boundary, we find that current benchmarks are adequate for high-frequency harm categories and several orders of magnitude short for rare, catastrophic ones. Safety benchmarks are not uninformative. They are informative about a specific and computable set of propositions, and the discipline they need is to state which.

Large Language Models (LLMs) are increasingly vulnerable to sophisticated multi-turn manipulation attacks, where adversaries strategically build context through seemingly benign conversational turns to circumvent safety measures and elicit harmful or unauthorized responses. These attacks exploit the temporal nature of dialogue to evade single-turn detection methods, representing a critical security vulnerability with significant implications for real-world deployments. This paper introduces the Temporal Context Awareness (TCA) framework, a novel defense mechanism designed to address this challenge by continuously analyzing semantic drift, cross-turn intention consistency and evolving conversational patterns. The TCA framework integrates dynamic context embedding analysis, cross-turn consistency verification, and progressive risk scoring to detect and mitigate manipulation attempts effectively. Preliminary evaluations on simulated adversarial scenarios demonstrate the framework's potential to identify subtle manipulation patterns often missed by traditional detection techniques, offering a much-needed layer of security for conversational AI systems. In addition to outlining the design of TCA , we analyze diverse attack vectors and their progression across multi-turn conversation, providing valuable insights into adversarial tactics and their impact on LLM vulnerabilities. Our findings underscore the pressing need for robust, context-aware defenses in conversational AI systems and highlight TCA framework as a promising direction for securing LLMs while preserving their utility in legitimate applications. We make our implementation available to support further research in this emerging area of AI security.

Jailbreak attacks against multimodal large language Models (MLLMs) are a significant research focus. Current research predominantly focuses on maximizing attack success rate (ASR), often overlooking whether the generated responses actually fulfill the attacker's malicious intent. This oversight frequently leads to low-quality outputs that bypass safety filters but lack substantial harmful content. To address this gap, we propose JPS, Jailbreak MLLMs with collaborative visual Perturbation and textual Steering, which achieves jailbreaks via corporation of visual image and textually steering prompt. Specifically, JPS utilizes target-guided adversarial image perturbations for effective safety bypass, complemented by "steering prompt" optimized via a multi-agent system to specifically guide LLM responses fulfilling the attackers' intent. These visual and textual components undergo iterative co-optimization for enhanced performance. To evaluate the quality of attack outcomes, we propose the Malicious Intent Fulfillment Rate (MIFR) metric, assessed using a Reasoning-LLM-based evaluator. Our experiments show JPS sets a new state-of-the-art in both ASR and MIFR across various MLLMs and benchmarks, with analyses confirming its efficacy. Codes are available at https://github.com/thu-coai/JPS{https://github.com/thu-coai/JPS}. warningcolor{Warning: This paper contains potentially sensitive contents.}