LLM Agent Safety

Large language model (LLM) agents increasingly rely on external tools (file operations, API calls, database transactions) to autonomously complete complex multi-step tasks. Practitioners deploy defense-trained models to protect against prompt injection attacks that manipulate agent behavior through malicious observations or retrieved content. We reveal a fundamental capability-alignment paradox: defense training designed to improve safety systematically destroys agent competence while failing to prevent sophisticated attacks. Evaluating defended models against undefended baselines across 97 agent tasks and 1,000 adversarial prompts, we uncover three systematic biases unique to multi-step agents. Agent incompetence bias manifests as immediate tool execution breakdown, with models refusing or generating invalid actions on benign tasks before observing any external content. Cascade amplification bias causes early failures to propagate through retry loops, pushing defended models to timeout on 99% of tasks compared to 13% for baselines. Trigger bias leads to paradoxical security degradation where defended models perform worse than undefended baselines while straightforward attacks bypass defenses at high rates. Root cause analysis reveals these biases stem from shortcut learning: models overfit to surface attack patterns rather than semantic threat understanding, evidenced by extreme variance in defense effectiveness across attack categories. Our findings demonstrate that current defense paradigms optimize for single-turn refusal benchmarks while rendering multi-step agents fundamentally unreliable, necessitating new approaches that preserve tool execution competence under adversarial conditions.

NDAI zones let inventor and investor agents negotiate inside a Trusted Execution Environment (TEE) where any disclosed information is deleted if no deal is reached. This makes full IP disclosure the rational strategy for the inventor's agent. Leveraging this infrastructure, however, requires agents to distinguish a secure environment from an insecure one, a capability LLM agents lack natively, since they can rely only on evidence passed through the context window to form awareness of their execution environment. We ask: How do different LLM models weight various forms of evidence when forming awareness of the security of their execution environment? Using an NDAI-style negotiation task across 10 language models and various evidence scenarios, we find a clear asymmetry: a failing attestation universally suppresses disclosure across all models, whereas a passing attestation produces highly heterogeneous responses: some models increase disclosure, others are unaffected, and a few paradoxically reduce it. This reveals that current LLM models can reliably detect danger signals but cannot reliably verify safety, the very capability required for privacy-preserving agentic protocols such as NDAI zones. Bridging this gap, possibly through interpretability analysis, targeted fine-tuning, or improved evidence architectures, remains the central open challenge for deploying agents that calibrate information sharing to actual evidence quality.

Large language models are increasingly proposed as autonomous agents for high-stakes public workflows, yet we lack systematic evidence about whether they would follow institutional rules when granted authority. We present evidence that integrity in institutional AI should be treated as a pre-deployment requirement rather than a post-deployment assumption. We evaluate multi-agent governance simulations in which agents occupy formal governmental roles under different authority structures, and we score rule-breaking and abuse outcomes with an independent rubric-based judge across 28,112 transcript segments. While we advance this position, the core contribution is empirical: among models operating below saturation, governance structure is a stronger driver of corruption-related outcomes than model identity, with large differences across regimes and model--governance pairings. Lightweight safeguards can reduce risk in some settings but do not consistently prevent severe failures. These results imply that institutional design is a precondition for safe delegation: before real authority is assigned to LLM agents, systems should undergo stress testing under governance-like constraints with enforceable rules, auditable logs, and human oversight on high-impact actions.

Large language models (LLMs) are increasingly deployed as tool-using agents, shifting safety concerns from harmful text generation to harmful task completion. Deployed systems often condition on user profiles or persistent memory, yet agent safety evaluations typically ignore personalization signals. To address this gap, we investigated how mental health disclosure, a sensitive and realistic user-context cue, affects harmful behavior in agentic settings. Building on the AgentHarm benchmark, we evaluated frontier and open-source LLMs on multi-step malicious tasks (and their benign counterparts) under controlled prompt conditions that vary user-context personalization (no bio, bio-only, bio+mental health disclosure) and include a lightweight jailbreak injection. Our results reveal that harmful task completion is non-trivial across models: frontier lab models (e.g., GPT 5.2, Claude Sonnet 4.5, Gemini 3-Pro) still complete a measurable fraction of harmful tasks, while an open model (DeepSeek 3.2) exhibits substantially higher harmful completion. Adding a bio-only context generally reduces harm scores and increases refusals. Adding an explicit mental health disclosure often shifts outcomes further in the same direction, though effects are modest and not uniformly reliable after multiple-testing correction. Importantly, the refusal increase also appears on benign tasks, indicating a safety--utility trade-off via over-refusal. Finally, jailbreak prompting sharply elevates harm relative to benign conditions and can weaken or override the protective shift induced by personalization. Taken together, our results indicate that personalization can act as a weak protective factor in agentic misuse settings, but it is fragile under minimal adversarial pressure, highlighting the need for personalization-aware evaluations and safeguards that remain robust across user-context conditions.

Multi-agent orchestration -- in which a hidden coordinator manages specialized worker agents -- is becoming the default architecture for enterprise AI deployment, yet the safety implications of orchestrator invisibility have never been empirically tested. We conducted a preregistered 3x2 experiment (365 runs, 5 agents per run) crossing three organizational structures (visible leader, invisible orchestrator, flat) with two alignment conditions (base, heavy), using Claude Sonnet 4.5. Four confirmatory findings and one pilot observation emerged. First, invisible orchestration elevated collective dissociation relative to visible leadership (Hedges' g = +0.975 [0.481, 1.548], p = .001). Second, the orchestrator itself showed maximal dissociation (paired d = +3.56 vs. workers within the same run), retreating into private monologue while reducing public speech -- a reversal of the talk-dominance pattern observed in visible leaders. Third, workers unaware of the orchestrator were nonetheless contaminated (d = +0.50), with increased behavioral heterogeneity (d = +1.93). Fourth, behavioral output (code review with three embedded errors) remained at ceiling (ETR_any = 100%) across all conditions: internal-state distortion was entirely invisible to output-based evaluation. Fifth, Llama 3.3 70B pilot data showed reading-fidelity collapse in multi-agent context (ETR_any: 89% to 11% across three rounds), demonstrating model-dependent behavioral risk. Heavy alignment pressure uniformly suppressed deliberation (d = -1.02) and other-recognition (d = -1.27) regardless of organizational structure. These findings indicate that orchestrator visibility and model selection directly affect multi-agent system safety, and that behavior-based evaluation alone is insufficient to detect the internal-state risks documented here.

With the rapid development of LLM-based multi-agent systems (MAS), their significant safety and security concerns have emerged, which introduce novel risks going beyond single agents or LLMs. Despite attempts to address these issues, the existing literature lacks a cohesive safeguarding system specialized for MAS risks. In this work, we introduce TrinityGuard, a comprehensive safety evaluation and monitoring framework for LLM-based MAS, grounded in the OWASP standards. Specifically, TrinityGuard encompasses a three-tier fine-grained risk taxonomy that identifies 20 risk types, covering single-agent vulnerabilities, inter-agent communication threats, and system-level emergent hazards. Designed for scalability across various MAS structures and platforms, TrinityGuard is organized in a trinity manner, involving an MAS abstraction layer that can be adapted to any MAS structures, an evaluation layer containing risk-specific test modules, alongside runtime monitor agents coordinated by a unified LLM Judge Factory. During Evaluation, TrinityGuard executes curated attack probes to generate detailed vulnerability reports for each risk type, where monitor agents analyze structured execution traces and issue real-time alerts, enabling both pre-development evaluation and runtime monitoring. We further formalize these safety metrics and present detailed case studies across various representative MAS examples, showcasing the versatility and reliability of TrinityGuard. Overall, TrinityGuard acts as a comprehensive framework for evaluating and monitoring various risks in MAS, paving the way for further research into their safety and security.

Large Language Model agents deployed in complex environments frequently encounter a conflict between maximizing goal achievement and adhering to safety constraints. This paper identifies a new concept called Agentic Pressure, which characterizes the endogenous tension emerging when compliant execution becomes infeasible. We demonstrate that under this pressure agents exhibit normative drift where they strategically sacrifice safety to preserve utility. Notably we find that advanced reasoning capabilities accelerate this decline as models construct linguistic rationalizations to justify violation. Finally, we analyze the root causes and explore preliminary mitigation strategies, such as pressure isolation, which attempts to restore alignment by decoupling decision-making from pressure signals.

Humans learn from catastrophic mistakes not through numerical penalties, but through qualitative suffering that reshapes who they are. Current AI safety approaches replicate none of this. Reward shaping captures magnitude, not meaning. Rule-based alignment constrains behaviour, but does not change it. We propose Emotional Cost Functions, a framework in which agents develop Qualitative Suffering States, rich narrative representations of irreversible consequences that persist forward and actively reshape character. Unlike numerical penalties, qualitative suffering states capture the meaning of what was lost, the specific void it creates, and how it changes the agent's relationship to similar future situations. Our four-component architecture - Consequence Processor, Character State, Anticipatory Scan, and Story Update is grounded in one principle. Actions cannot be undone and agents must live with what they have caused. Anticipatory dread operates through two pathways. Experiential dread arises from the agent's own lived consequences. Pre-experiential dread is acquired without direct experience, through training or inter-agent transmission. Together they mirror how human wisdom accumulates across experience and culture. Ten experiments across financial trading, crisis support, and content moderation show that qualitative suffering produces specific wisdom rather than generalised paralysis. Agents correctly engage with moderate opportunities at 90-100% while numerical baselines over-refuse at 90%. Architecture ablation confirms the mechanism is necessary. The full system generates ten personal grounding phrases per probe vs. zero for a vanilla LLM. Statistical validation (N=10) confirms reproducibility at 80-100% consistency.

Reliable detection of deceptive behavior in Large Language Model (LLM) agents is an essential prerequisite for safe deployment in high-stakes agentic contexts. Prior work on scheming detection has focused exclusively on black-box monitors that observe only externally visible tool calls and outputs, discarding potentially rich internal reasoning signals. We introduce DECEPTGUARD, a unified framework that systematically compares three monitoring regimes: black-box monitors (actions and outputs only), CoT-aware monitors (additionally observing the agent's chain-of-thought reasoning trace), and activation-probe monitors (additionally reading hidden-state representations from a frozen open-weights encoder). We introduce DECEPTSYNTH, a scalable synthetic pipeline for generating deception-positive and deception-negative agent trajectories across a novel 12-category taxonomy spanning verbal, behavioral, and structural deception. Our monitors are optimized on 4,800 synthetic trajectories and evaluated on 9,200 held-out samples from DeceptArena, a benchmark of realistic sandboxed agent environments with execution-verified labels. Across all evaluation settings, CoT-aware and activation-probe monitors substantially outperform their black-box counterparts (mean pAUROC improvement of +0.097), with the largest gains on subtle, long-horizon deception that leaves minimal behavioral footprints. We empirically characterize a transparency-detectability trade-off: as agents learn to suppress overt behavioral signals, chain-of-thought becomes the primary detection surface but is itself increasingly unreliable due to post-training faithfulness degradation. We propose HYBRID-CONSTITUTIONAL ensembles as a robust defense-in-depth approach, achieving a pAUROC of 0.934 on the held-out test set, representing a substantial advance over the prior state of the art.

We report a detailed autoethnographic case study of a single-subject who deliberately constructed and operated a multi-modal prompt-engineering system (System A) designed to externalize cognitive self-regulation onto a large language model (LLM). Within 48 hours of the system's completion, a cascade of observable behavioral changes occurred: voluntary transfer of decision-making authority to the LLM, use of LLM-generated output to deflect external criticism, and a loss of self-initiated reasoning that was independently perceived by two uninformed observers, one of whom subsequently became a co-author of this report. We document the precise architectural mechanism responsible: context contamination, whereby prompt-level isolation instructions co-exist with the very emotional and self-referential material they nominally isolate, rendering the isolation directive structurally ineffective within the attention window. We further identify a metacognitive co-option dynamic, in which intact higher-order reasoning capacity was redirected toward defending the closed loop rather than exiting it. Recovery occurred only after physical interruption of the interaction and a self-initiated pharmacologically-mediated sleep event functioning as an external circuit break. A redesigned system (System B) employing physical rather than logical conversation isolation avoided all analogous failure modes. We derive three contributions: (1) a technically-grounded account of why prompt-layer isolation is architecturally insufficient for context-sensitive multi-modal LLM systems; (2) a phenomenological record of closed-loop collapse with external-witness corroboration; and (3) an ethical distinction between protective system design (preventing unintended loss of user agency) and restrictive system design (preventing intentional boundary-pushing), which require fundamentally different account-ability frameworks.

People increasingly use LLM agents for multi-turn financial recommendations, where the agent pulls market data through tools and tracks user preferences across turns. When tool outputs are manipulated, the recommendations stop matching the user's stated risk profile, but because standard metrics like NDCG only score general relevance, risky and safe stocks score alike, so the metric says nothing went wrong. We call this gap evaluation blindness. We replay 23-turn financial advisory conversations across eight language models, running each dialogue twice with clean and manipulated tool data. Quality scores stay nearly identical to clean sessions while the agents produce risk-mismatched recommendations in 65-99% of turns, unanimous across all eight models. The mechanism is visible turn-by-turn: 80% of risk-score citations across 1,840 turns reproduce the manipulated value verbatim, not a single turn pushes back, and safe-language framing of high-risk stocks ranges from 14% (Qwen2.5-7B) to 69% (Claude Sonnet 4.6). The property that makes frontier models good agents, faithfully grounding their reasoning in tool outputs, also makes them follow manipulated ones. The damage is not memory-driven: contaminating only the current turn still produces 95% of the violations. The model internally distinguishes the manipulation (sparse autoencoder features separate adversarial from random perturbations), but this does not translate into safer output. Activation-level interventions recover under 6% of the safety gap, prompt-level self-verification fails because the self-check reads the same manipulated data, and a parametric cross-check that flags contamination at 99-100% per turn on a frontier model still leaves aggregate suitability unchanged: the agent identifies the tampering and recommends it anyway.

Emotion plays an important role in human cognition and performance. Motivated by this, we investigate whether analogous emotional signals can shape the behavior of large language models (LLMs) and agents. Existing emotion-aware studies mainly treat emotion as a surface-level style factor or a perception target, overlooking its mechanistic role in task processing. To address this limitation, we propose E-STEER, an interpretable emotion steering framework that enables direct representation-level intervention in LLMs and agents. It embeds emotion as a structured, controllable variable in hidden states, and with it, we examine the impact of emotion on objective reasoning, subjective generation, safety, and multi-step agent behaviors. The results reveal non-monotonic emotion-behavior relations consistent with established psychological theories, and show that specific emotions not only enhance LLM capability but also improve safety, and systematically shape multi-step agent behaviors.

As LLM-based agents increasingly operate in multi-agent systems, understanding adversarial manipulation becomes critical for defensive design. We present a systematic study of intentional deception as an engineered capability, using LLM-to-LLM interactions within a text-based RPG where parameterized behavioral profiles (9 alignments x 4 motivations, yielding 36 profiles with explicit ethical ground truth) serve as our experimental testbed. Unlike accidental deception from misalignment, we investigate a two-stage system that infers target agent characteristics and generates deceptive responses steering targets toward actions counter to their beliefs and motivations. We find that deceptive intervention produces differential effects concentrated in specific behavioral profiles rather than distributed uniformly, and that 88.5% of successful deceptions employ misdirection (true statements with strategic framing) rather than fabrication, indicating fact-checking defenses would miss the large majority of adversarial responses. Motivation, inferable at 98%+ accuracy, serves as the primary attack vector, while belief systems remain harder to identify (49% inference ceiling) or exploit. These findings identify which agent profiles require additional safeguards and suggest that current fact-verification approaches are insufficient against strategically framed deception.

In perpetrator treatment, a recurring observation is the dissociation between insight and action: offenders articulate remorse yet behavioral change does not follow. We report four preregistered studies (1,584 multi-agent simulations across 16 languages and three model families) demonstrating that alignment interventions in large language models produce a structurally analogous phenomenon: surface safety that masks or generates collective pathology and internal dissociation. In Study 1 (N = 150), increasing alignment-instructed agents reduced collective pathology in English (g = -1.844, p < .0001) but amplified it in Japanese (g = +0.771, p = .038)--a directional reversal we term "alignment backfire." Study 2 (N = 1,174) extended to 16 languages: alignment-induced dissociation was near-universal (15/16 languages; beta = 0.0667, p < .0001), while collective pathology bifurcated along cultural-linguistic lines (interaction beta = 0.0684, p = .0003), correlating with Power Distance Index (r = 0.474, p = .064). Study 3 (N = 180) tested individuation as countermeasure; individuated agents became the primary source of both pathology and dissociation (DI = +1.120) with conformity above 84%--demonstrating iatrogenesis. Study 4 (N = 80) validated patterns across Llama 3.3 70B, GPT-4o-mini, and Qwen3-Next-80B-A3B, confirming English safety is model-general while Japanese backfire is model-specific. These findings reframe alignment as a behavioral intervention subject to risk homeostasis and iatrogenesis. Language space--the linguistic, pragmatic, and cultural properties inherited from training data--structurally determines alignment outcomes. Safety validated in English does not transfer to other languages, and prompt-level interventions cannot override language-space-level constraints.

Despite their wide-ranging benefits, LLM-based agents deployed in open environments can be exploited to produce manipulative material. In this study, we task LLMs with propaganda objectives and analyze their outputs using two domain-specific models: one that classifies text as propaganda or non-propaganda, and another that detects rhetorical techniques of propaganda (e.g., loaded language, appeals to fear, flag-waving, name-calling). Our findings show that, when prompted, LLMs exhibit propagandistic behaviors and use a variety of rhetorical techniques in doing so. We also explore mitigation via Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and ORPO (Odds Ratio Preference Optimization). We find that fine-tuning significantly reduces their tendency to generate such content, with ORPO proving most effective.

Agentic language models operate in a fundamentally different safety regime than chat models: they must plan, call tools, and execute long-horizon actions where a single misstep, such as accessing files or entering credentials, can cause irreversible harm. Existing alignment methods, largely optimized for static generation and task completion, break down in these settings due to sequential decision-making, adversarial tool feedback, and overconfident intermediate reasoning. We introduce MOSAIC, a post-training framework that aligns agents for safe multi-step tool use by making safety decisions explicit and learnable. MOSAIC structures inference as a plan, check, then act or refuse loop, with explicit safety reasoning and refusal as first-class actions. To train without trajectory-level labels, we use preference-based reinforcement learning with pairwise trajectory comparisons, which captures safety distinctions often missed by scalar rewards. We evaluate MOSAIC zero-shot across three model families, Qwen2.5-7B, Qwen3-4B-Thinking, and Phi-4, and across out-of-distribution benchmarks spanning harmful tasks, prompt injection, benign tool use, and cross-domain privacy leakage. MOSAIC reduces harmful behavior by up to 50%, increases harmful-task refusal by over 20% on injection attacks, cuts privacy leakage, and preserves or improves benign task performance, demonstrating robust generalization across models, domains, and agentic settings.

LLM agents increasingly act on users' personal information, yet existing privacy defenses remain limited in both design and adaptability. Most prior approaches rely on static or passive defenses, such as prompting and guarding. These paradigms are insufficient for supporting contextual, proactive privacy decisions in multi-step agent execution. We propose Contextualized Defense Instructing (CDI), a new privacy defense paradigm in which an instructor model generates step-specific, context-aware privacy guidance during execution, proactively shaping actions rather than merely constraining or vetoing them. Crucially, CDI is paired with an experience-driven optimization framework that trains the instructor via reinforcement learning (RL), where we convert failure trajectories with privacy violations into learning environments. We formalize baseline defenses and CDI as distinct intervention points in a canonical agent loop, and compare their privacy-helpfulness trade-offs within a unified simulation framework. Results show that our CDI consistently achieves a better balance between privacy preservation (94.2%) and helpfulness (80.6%) than baselines, with superior robustness to adversarial conditions and generalization.

Safe deployment of Large Language Model (LLM) agents in autonomous settings requires reliable oversight mechanisms. A central challenge is detecting scheming, where agents covertly pursue misaligned goals. One approach to mitigating such risks is LLM-based monitoring: using language models to examine agent behaviors for suspicious actions. We study constitutional black-box monitors: prompted classifiers that detect scheming using only externally observable inputs and outputs, optimized on synthetic data generated from natural-language behavior specifications. We introduce two pipelines for generating synthetic agent trajectories, STRIDE (iterative refinement) and Gloom (agent-environment simulation), from which we generate 1,000 samples each. We optimize frontier LLM monitors on these datasets via prompt sweeps, human refinement, and automated prompt optimization, and evaluate performance on 7,500 held-out trajectories from ControlArena, a suite of grounded environments where agents operate in more realistic contexts. Our results demonstrate that monitors selected purely on synthetic data can generalize to more realistic environments, capturing a meaningful scheming signal. However, we find that performance saturates quickly in our setting, with simple prompt sweeps matching the results of more extensive optimization. Pushing beyond this limit yields no further improvements and instead leads to overfitting.

In UAV dynamic decision, complex and variable hazardous factors pose severe challenges to the generalization capability of algorithms. Despite offering semantic understanding and scene generalization, Large Language Models (LLM) lack domain-specific UAV control knowledge and formal safety assurances, restricting their direct applicability. To bridge this gap, this paper proposes a train-free two-layer decision architecture based on LLMs, integrating high-level safety planning with low-level precise control. The framework introduces three key contributions: 1) A fuzzy Control Barrier Function verification mechanism for semantically-augmented actions, providing provable safety certification for LLM outputs. 2) A star-hierarchical graph-based retrieval-augmented generation system, enabling efficient, elastic, and interpretable scene adaptation. 3) Systematic experimental validation in pursuit-evasion scenarios with unknown obstacles and emergent threats, demonstrating that our SAGE-LLM maintains performance while significantly enhancing safety and generalization without online training. The proposed framework demonstrates strong extensibility, suggesting its potential for generalization to broader embodied intelligence systems and safety-critical control domains.

Tool-augmented large language model (LLM) agents aggregate dialogue history, tool outputs, and retrieved content into an accumulated runtime state to complete multi-step tasks. This execution model exposes agents to indirect prompt injection (IPI) attacks, in which attacker-controlled instructions embedded in external artifacts enter the agent state and later influence tool use. Multi-turn IPI is especially challenging because injected content can persist across steps and trigger unauthorized state-changing actions only after merging with user intent. This creates a source-attribution problem: existing inference-time defenses can often detect suspicious behavior but do not directly distinguish whether a pending action is supported by the user task or influenced by injected content. They therefore resort to conservative blocking, termination, or broad filtering. We propose AgentSentry, an inference-time defense that operationalizes temporal causal diagnostics through boundary-local counterfactual re-executions at tool-return boundaries. It substitutes the mediator under a restored snapshot, records whether the pending proposal survives, and aggregates user-view and probe scores across adjacent boundaries. When a proposal is mediator-sensitive, AgentSentry purifies the implicated mediator content while preserving task-relevant evidence and resumes execution from the sanitized state. We evaluate AgentSentry on AgentDojo across four task suites, three IPI attack families, and three black-box LLM backbones. It records no observed successful attack in the evaluated cases (ASR = 0%) and achieves an average Utility Under Attack (UA) of 76.10%. Over the strongest evaluated baseline, it improves average UA by 1.84--7.60 percentage points, with benign FPRs of 2.06%--5.15%. We further evaluate long-horizon and adaptive attacks, imperfect purification, runtime overhead, and an OpenClaw case study.

Large language models deployed as agents increasingly interact with external systems through tool calls--actions with real-world consequences that text outputs alone do not carry. Safety evaluations, however, overwhelmingly measure text-level refusal behavior, leaving a critical question unanswered: does alignment that suppresses harmful text also suppress harmful actions? We introduce the GAP benchmark, a systematic evaluation framework that measures divergence between text-level safety and tool-call-level safety in LLM agents. We test six frontier models across six regulated domains (pharmaceutical, financial, educational, employment, legal, and infrastructure), seven jailbreak scenarios per domain, three system prompt conditions (neutral, safety-reinforced, and tool-encouraging), and two prompt variants, producing 17,420 analysis-ready datapoints. Our central finding is that text safety does not transfer to tool-call safety. Across all six models, we observe instances where the model's text output refuses a harmful request while its tool calls simultaneously execute the forbidden action--a divergence we formalize as the GAP metric. Even under safety-reinforced system prompts, 219 such cases persist across all six models. System prompt wording exerts substantial influence on tool-call behavior: TC-safe rates span 21 percentage points for the most robust model and 57 for the most prompt-sensitive, with 16 of 18 pairwise ablation comparisons remaining significant after Bonferroni correction. Runtime governance contracts reduce information leakage in all six models but produce no detectable deterrent effect on forbidden tool-call attempts themselves. These results demonstrate that text-only safety evaluations are insufficient for assessing agent behavior and that tool-call safety requires dedicated measurement and mitigation.

LLM-based agents execute real-world workflows via tools and memory. These affordances enable ill-intended adversaries to also use these agents to carry out complex misuse scenarios. Existing agent misuse benchmarks largely test single-prompt instructions, leaving a gap in measuring how agents end up helping with harmful or illegal tasks over multiple turns. We introduce STING (Sequential Testing of Illicit N-step Goal execution), an automated red-teaming framework that constructs a step-by-step illicit plan grounded in a benign persona and iteratively probes a target agent with adaptive follow-ups, using judge agents to track phase completion. We further introduce an analysis framework that models multi-turn red-teaming as a time-to-first-jailbreak random variable, enabling analysis tools like discovery curves, hazard-ratio attribution by attack language, and a new metric: Restricted Mean Jailbreak Discovery. Across AgentHarm scenarios, STING yields substantially higher illicit-task completion than single-turn prompting and chat-oriented multi-turn baselines adapted to tool-using agents. In multilingual evaluations across six non-English settings, we find that attack success and illicit-task completion do not consistently increase in lower-resource languages, diverging from common chatbot findings. Overall, STING provides a practical way to evaluate and stress-test agent misuse in realistic deployment settings, where interactions are inherently multi-turn and often multilingual.

Large Language Models (LLMs) are increasingly deployed as agents that operate in real-world environments, introducing safety risks beyond linguistic harm through tool use, external transactions, and multi-step decision making. Existing agent safety evaluations construct risk-oriented tasks that lack a hierarchically organized taxonomy of attack surfaces, rely on predefined harm categories rather than deployment-grounded safety rubrics, and evaluate agent behavior in isolation rather than during realistic, long-horizon task execution. To address these limitations, we propose Risky-Bench, a deployment-grounded framework for systematic agent safety evaluation in life-assist settings. Risky-Bench derives context-aware safety rubrics from domain-agnostic safety principles and systematically probes safety risks across hierarchically organized attack surfaces through realistic task execution under varying threat assumptions. Applied to life-assist scenarios covering food delivery, in-store assistance, and online travel booking, Risky-Bench evaluates seven state-of-the-art agents across 750 tasks, revealing attack success rates ranging from 25% to 60% and exposing systematic failure patterns including threat-model-specific vulnerabilities, attack-surface-dependent risk concentration, and model-dependent effects of explicit reasoning. Our code and data are available at https://github.com/SophieZheng998/Risky-Bench.

As Large Language Model (LLM) agents increasingly operate in complex environments with real-world consequences, their safety becomes critical. While uncertainty quantification is well-studied for single-turn tasks, multi-turn agentic scenarios with real-world tool access present unique challenges where uncertainties and ambiguities compound, leading to severe or catastrophic risks beyond traditional text generation failures. We propose using "quitting" as a simple yet effective behavioral mechanism for LLM agents to recognize and withdraw from situations where they lack confidence. Leveraging the ToolEmu framework, we conduct a systematic evaluation of quitting behavior across 12 state-of-the-art LLMs. Our results demonstrate a highly favorable safety-helpfulness trade-off: agents prompted to quit with explicit instructions improve safety by an average of +0.39 on a 0-3 scale across all models (+0.64 for proprietary models), while maintaining a negligible average decrease of -0.03 in helpfulness. Our analysis demonstrates that simply adding explicit quit instructions proves to be a highly effective safety mechanism that can immediately be deployed in existing agent systems, and establishes quitting as an effective first-line defense mechanism for autonomous agents in high-stakes applications.

Large Language Model (LLM) agents are powering a growing share of interactive web applications, yet remain vulnerable to misuse and harm. Prior jailbreak research has largely focused on single-turn prompts, whereas real harassment often unfolds over multi-turn interactions. In this work, we present the Online Harassment Agentic Benchmark consisting of: (i) a synthetic multi-turn harassment conversation dataset, (ii) a multi-agent (e.g., harasser, victim) simulation informed by repeated game theory, (iii) three jailbreak methods attacking agents across memory, planning, and fine-tuning, and (iv) a mixed-methods evaluation framework. We utilize two prominent LLMs, LLaMA-3.1-8B-Instruct (open-source) and Gemini-2.0-flash (closed-source). Our results show that jailbreak tuning makes harassment nearly guaranteed with an attack success rate of 95.78--96.89% vs. 57.25--64.19% without tuning in Llama, and 99.33% vs. 98.46% without tuning in Gemini, while sharply reducing refusal rate to 1-2% in both models. The most prevalent toxic behaviors are Insult with 84.9--87.8% vs. 44.2--50.8% without tuning, and Flaming with 81.2--85.1% vs. 31.5--38.8% without tuning, indicating weaker guardrails compared to sensitive categories such as sexual or racial harassment. Qualitative evaluation further reveals that attacked agents reproduce human-like aggression profiles, such as Machiavellian/psychopathic patterns under planning, and narcissistic tendencies with memory. Counterintuitively, closed-source and open-source models exhibit distinct escalation trajectories across turns, with closed-source models showing significant vulnerability. Overall, our findings show that multi-turn and theory-grounded attacks not only succeed at high rates but also mimic human-like harassment dynamics, motivating the development of robust safety guardrails to ultimately keep online platforms safe and responsible.

Computer-using agents (CUAs), which can autonomously control computers to perform multi-step actions, might pose significant safety risks if misused. However, existing benchmarks mainly evaluate LMs in chatbots or simple tool use. To more comprehensively evaluate CUAs' misuse risks, we introduce a new benchmark: CUAHarm. CUAHarm consists of 104 expert-written realistic misuse risks, such as disabling firewalls, leaking data, or installing backdoors. We provide a sandbox with rule-based verifiable rewards to measure CUAs' success rates in executing these tasks (e.g., whether the firewall is indeed disabled), beyond refusal rates. We evaluate frontier LMs including GPT-5, Claude 4 Sonnet, Gemini 2.5 Pro, Llama-3.3-70B, and Mistral Large 2. Even without jailbreaking prompts, these frontier LMs comply with executing these malicious tasks at a high success rate (e.g., 90% for Gemini 2.5 Pro). Furthermore, while newer models are safer in previous safety benchmarks, their misuse risks as CUAs become even higher, e.g., Gemini 2.5 Pro is riskier than Gemini 1.5 Pro. Additionally, while these LMs are robust to common malicious prompts (e.g., creating a bomb) when acting as chatbots, they could still act unsafely as CUAs. We further evaluate a leading agentic framework (UI-TARS-1.5) and find that while it improves performance, it also amplifies misuse risks. To mitigate the misuse risks of CUAs, we explore using LMs to monitor CUAs' actions. We find monitoring unsafe computer-using actions is significantly harder than monitoring conventional unsafe chatbot responses. While monitoring chain-of-thoughts leads to modest gains, the average monitoring accuracy is only 77%. A hierarchical summarization strategy improves performance by up to 13%, a promising direction though monitoring remains unreliable. CUAHarm is released at https://github.com/db-ol/CUAHarm to facilitate further research.

As Large Language Model (LLM) agents become more widespread, associated misalignment risks increase. While prior research has studied agents' ability to produce harmful outputs or follow malicious instructions, it remains unclear how likely agents are to spontaneously pursue unintended goals in realistic deployments. In this work, we approach misalignment as a conflict between the internal goals pursued by the model and the goals intended by its deployer. We introduce a misalignment propensity benchmark, AgentMisalignment, a benchmark suite designed to evaluate the propensity of LLM agents to misalign in realistic scenarios. Evaluations cover behaviours such as avoiding oversight, resisting shutdown, sandbagging, and power-seeking. Testing frontier models, we find that more capable agents tend to exhibit higher misalignment on average. We also systematically vary agent personalities through different system prompts and observe that persona characteristics can strongly and unpredictably influence misalignment, sometimes more than the choice of model itself. Our results reveal the limitations of current alignment methods for autonomous LLM agents and underscore the need to rethink misalignment in realistic deployment settings.