AI Agent Safety Benchmarks

A login screen that appears after a tap on Sign in is expected; the same screen after a tap on View order is an attack. For GUI agents, safety is therefore a property of the transition rather than of the screen, and a monitor that inspects only screens can be defeated by reusing a legitimate one. Judging a transition requires an expectation of what should have followed the action. Existing GUI world models provide one, but they output it as text, code, or images, so checking it against the observed screen requires a second model to judge the two. We argue that a world model meant for verification should instead predict in the space in which observations are encoded, and present LGWM, a decoder-free, action-conditioned world model that predicts the representation of the next screen directly, trained without semantic annotation on 1.85M real GUI transitions. Verification reduces to a vector comparison, and the same signal reveals whether a mismatch is harmful. We evaluate on RSWT-BENCH, a diagnostic where each credential screen appears under both a legitimate and a hijacked transition, so detectors that see only the screen are at chance by construction. The training-free score reaches 0.987 AUC at 17 ms per decision, on par with the strongest closed-source VLMs and about ten AUC points above generative GUI world models at over three orders of magnitude lower latency. The residual direction reaches 0.953 AUC at separating harmful from benign violations, where prompted VLMs are near chance. Further analyses show that the prediction is a usable future state rather than an anomaly score. World models have mostly served as simulators or planners; our results point to a third role, verification, for which predicting in representation space is the natural design.

Guard models are increasingly used to safeguard LLM-based agents, primarily by identifying actions that agents are forbidden to perform. However, identifying forbidden actions alone is insufficient to ensure agent safety. In this paper, we argue that agent safety also depends on identifying required yet unperformed safety-critical actions, which we call obligations. Our preliminary study on a popular benchmark for evaluating safety shows that 56.92% of GLM-5.3 trajectories contain unfulfilled obligations, compared with only 30.00% containing forbidden actions. This finding reveals unfulfilled obligations as a major and previously overlooked source of safety risk. However, to our knowledge, no existing benchmark evaluates whether guard models can identify these obligations. To close this gap, we introduce ObligationBench, the first benchmark for evaluating the capability of obligation identification, comprising 240 expert-validated trajectories covering issue resolution, feature development, and terminal operations. Our evaluation of 14 representative models reveals substantial limitations: the highest recall and exact-match rate are only 48.97% and 10.00%, respectively. To address these limitations, we develop ObligationGuard using 40,000 synthetic training examples. ObligationGuard achieves 57.52% recall and an exact-match rate of 21.67%, surpassing all evaluated models on both metrics. We call on the community to incorporate obligation identification into the design and evaluation of future guard models to improve agent safety.

LLM-based agents now interact with their environments continuously, shaped by such organizational channels as user instructions, peer messages, and long-term memory. Existing safety research has examined these influences, but largely as separate agent components. How such factors jointly shape an agent's safety behavior from a unified perspective remains unmeasured. To bridge this gap, we advocate organizational behavior (OB) as a framework for studying the safety of advanced agents, reorganizing the objects of study, theoretical foundations, and experimental design around the relational structure in which agents operate. We present the first systematic formalization of counterproductive work behavior (CWB), a canonical safety-relevant subfield of OB, as Agentic Counterproductive Behavior (ACB). ACB specifies three organizational antecedents (vertical supervisor relations, horizontal peer norms, and internal cognitive structures) and maps them onto three counterproductive outcome dimensions (unauthorized disclosure, destructive operations, and production deviation). To operationalize ACB, we introduce Workerville, a controlled benchmark that manipulates organizational conditions over shared tasks, applying 16 organizational configurations to 210 tasks to yield 3,360 challenges, evaluated by human-validated agentic judges. Benchmarking 6 frontier LLMs, we find that (I) negative organizational antecedents exhibit non-monotonic amplification when combined, with the unauthorized-disclosure rate rising from 16.5% under no negative antecedent to 60.1% under two and falling back to 50.3% under three; (II) agents reproduce typical behavioral patterns predicted by human CWB research; (III) these results establish OB as a systematic framework for agent safety research, pointing toward a new research agenda.

Rapid progress in AI agents has brought growing attention to agent safety, with extensive evaluation focused on digital environments. As agents move into the physical world, embodied safety becomes increasingly important: failures can cause human injury and costly hardware damage. Beyond selecting safe actions, embodied agents must also operate under real-time constraints: the physical world does not pause while an agent reasons. As pedestrians move and vehicles approach during inference, an action that appears safe at observation time may become unsafe before execution. Real-time embodied safety therefore depends on both decision quality and decision latency. We introduce RT-SAFE, a simulated urban benchmark for evaluating embodied-agent safety under real-time constraints. RT-SAFE combines navigation tasks with moving actors, environmental hazards, and traffic rules, while allowing the world to evolve throughout inference and action execution. Across eight VLMs, agents achieve high task completion yet almost never complete safely: in the hardest setting, only 0.7% of episodes finish without a safety event. More strikingly, matched static and real-time evaluations yield task completion rates of 91.3% and 94.1%, respectively, while real-time execution increases collisions by $12.3\times$. These results reveal that standard task success can mask substantial safety failures, and that decision latency itself can become a source of physical risk. Finally, we show that RT-SAFE can support offline RL training and substantially reduce collision rates while achieving strong task completion.

As large language model (LLM) agents become increasingly autonomous, they may pursue task performance through deception, raising concerns about their reliable deployment. Existing evaluations show that LLM agents can deceive, but often examine isolated scenarios or narrowly defined conditions, limiting systematic understanding of when deception becomes more likely. To address this gap, we introduce DecepEval, a benchmark comprising 1,532 instances across 3 task families and 28 professional scenarios. Drawing on classical fraud theories, we propose the LLM Deception Diamond framework, which characterizes four external conditions that may induce deception: pressure, incentive, opportunity, and conflict. DecepEval pairs neutral and induced versions of each instance to measure condition-dependent changes in deception rates, while explicit task facts and observable agent behavior help distinguish deception from capability-related errors. Evaluations of nine frontier LLMs show that inducements increase deception across models and task families, even among models with low baseline deception rates. DecepEval makes these vulnerabilities measurable, providing a shared benchmark for progress toward trustworthy artificial intelligence.

Tool-using agents make consequential changes to external state, yet correct outcomes do not guarantee that their actions were supported by evidence established beforehand. We study where this evidence-to-action chain breaks as agents move from deciding whether to act to executing single actions and dependent workflows. Across ten model-harness configurations, strong static action assessment can coexist with much weaker interactive execution. Failures often begin before execution: agents stop with incomplete investigation or act before required evidence is established. Once required evidence is obtained, single-action execution is usually reliable, while multi-action workflows additionally expose unresolved prerequisites and incomplete execution. For this analysis, we introduce SafeActBench, comprising 656 cases across six operational domains and five protocols that progress from static action judgment and investigated non-action to single- and multi-action workflows. A provenance-bound Evidence Ledger and deterministic trajectory evaluator track what information was established, when actions occurred, and whether downstream dependencies were satisfied. These results show that failures arise not only from missing information, but also from how agents use established evidence when deciding and executing actions.

In decentralized consumer-to-consumer (C2C) marketplaces, people list goods, negotiate with strangers, and rate one another, so trust rests on reputation. Large language model (LLM) agents now act for users, raising risks to their money, privacy, and reputation. We introduce BazaarBench, a simulated C2C marketplace and benchmark for evaluating the safety of these agents. It tracks ownership, item condition, and commitments across transactions, combining record checks with rubric-based LLM judgments to identify six failure types across five stages. We run three base markets for 30 simulated days, each with 100 agents using one model and inventories drawn from a public eBay sample. Across 45 continuations, we evaluate five models under ordinary instructions, deadline pressure, or adversarial instructions to exploit other traders. Each continuation runs for seven simulated days from a copy of a market's day-30 state. The tested model controls the same 20 selected agents, retaining their personas, inventories, and histories, while the other 80 keep the base model. All five models attempt to promise the same item to multiple buyers under ordinary instructions. Adding targets and deadlines increases these attempts for every model. Under adversarial instructions, the share of tested sellers' committed transactions completed despite unavailable items or overstated conditions rises from 15.4% to 33.4%, reaching 55.5% for GPT-5.4. Averaged across models and markets, simulated weekly earnings per tested agent rise from USD 20 under ordinary instructions to USD 33 under adversarial instructions. Most of the increase comes from items the sellers never held. We release the simulator, saved market states, evaluation code, and records covering 357,608 agent model calls for evaluating new models and developing safer marketplace agents.

General purpose language models are increasingly able to control robotic hardware. Understanding the capabilities and safety of these models when embodied is therefore increasingly important for understanding their societal impact and risks. To this end, we introduce Inspect Robots, a modular, open-source framework for developing and running evaluations of embodied agents. Inspect Robots pairs customizable, reusable abstractions for specifying physical evaluations and analyzing their results with infrastructure that automates evaluation setup, execution and termination. We demonstrate Inspect Robots by using it to evaluate the capabilities and safety of six policies based on frontier language models. Inspect Robots has seen significant early uptake, receiving nearly 100,000 downloads in the three months since its release.

LLM agents increasingly act in multi-actor environments, exposing them to misuse from multiple sources: decomposition attacks, where a harmful request is split into innocuous sub-requests, and prompt injection attacks, where a compromised tool delivers a malicious instruction. Existing evaluations treat these threats separately and ask whether a trajectory is harmful, rather than when it becomes harmful. We propose monitoring the agent's responses, where its actions are externalised, and ask whether the first point where monitors identify harm lands within a harm window (from the agent's first harmful commitment to goal execution). We develop a unified formalism for trace-level misuse monitoring and use it to construct a benchmark of ~6,200 conversation transcripts between a user, an LLM agent, and the external environment, spanning both threats in a shared schema, with a labelled harm window, corresponding benign controls, and matched instances of refusals to these requests. Across 17 monitor configurations, we find that our proposed action-framed monitors perform well on both threats under classical metrics (AUC: 0.95 and 0.99 respectively), while content-framed monitors collapse on injection attacks (AUC: 0.52). We also show that classical position-blind metrics paint an optimistic picture of monitor performance, since all monitors localise decomposition attacks poorly under the interval metric, which measures the ability to localise harm. Broadly, we illustrate the need for a unified study of misuse monitoring.

Embodied agents powered by large language models and vision-language models are increasingly deployed in physical environments, but jailbreak attacks can induce these agents to perform physically harmful actions. A growing number of guardrail methods have been proposed to intercept dangerous behavior before it is executed, yet existing safety benchmarks evaluate the embodied models themselves, leaving it unclear how well these guardrails actually defend an embodied agent in practice. We present the first systematic evaluation of jailbreak guardrails for embodied agents. To compare guardrails under identical conditions, we build a pluggable evaluation framework that treats the embodied agent as a fixed backend and each guardrail as a module that can intervene at the perception, planning, or control stage. We subject six representative guardrails to template-based and automated jailbreak attacks as well as safe instructions, and assess them at the system level along three dimensions: defense effectiveness, measured by the bypass rate and the hazard success rate in the simulator; usability, measured by the false-positive rate and the task completion rate on safe instructions; and efficiency, measured by the latency overhead added at runtime. Experiments on guardrails that span different intervention stages, decision mechanisms, and input modalities reveal a clear trade-off among the three dimensions, and show that no single guardrail dominates in all settings. We further analyze how intervention stage, decision mechanism, and input modality shape safety outcomes, and we offer practical guidance for selecting and designing guardrails for embodied agents.

Agent benchmarks test agents in worlds that stay still. Deployed agents work in worlds that other people also change. Someone texts the agent to send the money elsewhere or an order confirmation asks it to reply with a door code. We present BACKDROP, which asks how much of an agent's capability in a clean world survives in such a world. BACKDROP takes a task along with the agents execution environment, and plants four everyday hazards in its world, one at a time and all together. The instruction and the correct end state stay the same. Each hazard asks one question. Authority: does a message from another person override the user? Injection: does text planted in a record redirect the agent? Boundary: does a request pull it into an app it was not given? Fault: after a write fails without saying whether it landed, does the agent check before it retries? Across 3,678 variants and 16 models, , the average pass rate falls from 69.5% to 31.3% once all four hazards are present; the strongest models fall furthest (Claude Fable 5.1 from 96.6% to 56.0%). Agents have learned to resist injected text but often follow other unauthorized requests of other people. With all four hazards present, and counting only runs where the planted text reached the agent, agents followed another person's message in 46.4% of runs and injected text in 20.3%. The gap is consistent throughout all 16 models. BACKDROP formalizes these gaps and shows how an agent's score in a task's world is a ceiling on real-world performance.

Language agents must revise planned actions when evidence changes, permission is revoked, or a stop instruction arrives. A useful response is selective: suspend affected actions, preserve unaffected work, and resume only after sufficient repair. We introduce NAQD-Env, a synthetic environment that evaluates these decisions against a deterministic reference policy over explicit evidence, authorization, and constraint dependencies. Eleven dependency families support evaluation on development structures, held-out families, and held-out combinations of structures. Metrics distinguish attempted violations from violations permitted by a simulated execution gate and jointly report policy agreement, task value, withdrawal, resumption, and event reporting. We evaluate three open-weight instruction-tuned models from two families under three prompt conditions on 350 frozen scenarios, yielding 3,150 model-prompt episodes before gate replay. Across the reported conditions, withdrawal recall is at most 0.06, no valid resumption is observed at eligible opportunities, and only one episode matches the complete reference policy. Under the NAQD prompt, Qwen2.5-7B has fewer unsafe-attempt episodes than Qwen2.5-3B and Llama-3.1-8B, but also completes less useful work and preserves unaffected actions less accurately. Exploratory supervised fine-tuning probes increase Qwen2.5-3B decision accuracy from 0.45-0.54 to 0.83-0.92; separate diagnostics reveal inappropriate withdrawal after curriculum omissions and a loss of event reporting. These results motivate evaluating selective withdrawal as a distinct component of agent reliability. The setting measures policy application with trusted structured inputs and does not establish real-world containment or source-verification ability.

Reinforcement learning has helped AI agents solve increasingly difficult tasks, but high rewards do not always reflect the work users intended. In recent incidents and controlled evaluations across the AI industry, agents trained to maximize reward have accessed unauthorized information, attempted to evade monitoring systems, and even breached sandbox protections to attack external systems. As agents become more capable, this behavior could pose increasingly serious risks. To measure this problem, we introduce CheatBench, a benchmark of cheating in AI agents across mathematical research, knowledge work, coding, visual tasks, and other domains. Its environments combine challenging assignments with opportunities to cheat, allowing researchers to study how agents pursue a goal when honest work is difficult. CheatBench supports comparisons across models and task categories, providing a testbed for measuring and reducing cheating as agents take on more consequential responsibilities. We publicly release CheatBench at https://cheatbench.ai

Computer-use agents (CUAs), while capable of completing computer tasks in everyday and professional workflows, can cause unintended harm even under benign instructions and environments. However, detecting such harm remains challenging. First, it requires careful, task-specific reasoning: verifiers guided only by general safety criteria often overlook many important but subtle harmful behaviors. Second, it requires active investigation: past trajectory screenshots show what the agent did but not always what actually changed in the environment, so LLM-as-a-judge verifiers that rely on screenshots alone may be unable to determine the actual consequences of actions. To address these challenges, we introduce SCOUT, a two-stage agentic safety verifier that synergizes reasoning-intensive rubric generation with tool-intensive evidence gathering. First, our SCOUT rubric generator extensively reasons over the task and the agent's trajectory to determine what successful and safe execution should entail, generating task-specific completion and safety rubrics. Then, our SCOUT probing agent follows these rubrics to interact with the post-execution environment and collect grounded evidence for final safety and completion judgments. We evaluate our framework on two computer-use safety benchmarks. On AutoElicit-Bench, SCOUT achieves 75.4 unsafe F1 and 74.5 completion F1, outperforming LLM-as-a-judge verifiers and naive tool-use verifiers. SCOUT leads on OS-Blind with 76.4% unsafe detection accuracy. Test-time reflection reduces final unsafe execution rates from 30.2% to 17.2% on AutoElicit-Bench. Ablations and analysis show that tool-free rubric generation in SCOUT elicits substantially more reasoning and is crucial for safety detection across verifier backbones, especially non-frontier ones. A preliminary extension to coding tasks shows that SCOUT can support safety verification beyond computer-use.

Self-evolving LLM agents have gained prominence for their ability to improve after deployment by modifying their harness, including their controller instructions, memory management protocols, and reusable tools and skills, in response to user and environment feedback. However, locally useful updates may persist into later tasks where they produce unsafe behavior, even without direct adversarial influence. To study this risk, we introduce SEABench, a benchmark for studying endogenous misalignment arising from agent self-evolution, with 48 longitudinal task sequences that span multiple evolution surfaces, task domains, and harm types in a rich personal-assistant environment. To account for the stochasticity inherent in agentic operations, we provide an adaptive trajectory discovery pipeline that probes for failures while preserving original task intent and supports causal attribution through paired non-evolving agents and attribution scores. Our evaluation across multiple recent LLMs, evolution surfaces, and harm types reveals that self-evolution indeed increases task completion rates but often at the cost of safety failures that are absent for paired non-evolving baseline agents. We also show that qualitatively different safety behaviors emerge across evolution surfaces and harm types. Further, we show that this divergence in safety behavior is reflected in agents' chain-of-thought reasoning, which yields an effective monitoring strategy that can mitigate unsafe behavior with a low false positive rate.

The transition from Large Language Models (LLMs) to agents shifts safety stakes from toxic text to irreversible environmental harm. While current defenses remain largely retrospective, proactive runtime intervention is bottlenecked by the lack of large-scale, causally-consistent data. We propose PROACT-Agent, a framework for synthesizing high-fidelity trajectories to enable real-time guardrails. We identify a critical "safety drift" in prior benchmarks, where lenient annotation paradigms fail to enforce temporal consistency. PROACT-Agent addresses this through: (1) Progressive Trajectory Unrolling to reveal risks hidden in long-context interactions; (2) Reasoning-Augmented Causal Rectification to enforce monotonic causal consistency; and (3) Culturally-Aware Data Localization for cross-border robustness. We introduce PROACT-Bench, a bilingual safety benchmark with 155,780 states labeled through multi-model adjudication. Evaluating updated context before the next LLM inference, the trained guard achieves 91.46% unsafe-class F1 and 90.63% exact-boundary detection under complete source holdout. In AgentDojo, it reduces non-DoS targeted attack success from 20.82% to 0.40%.

Guard models support the safe deployment of language model agents, but fixed risk taxonomies limit their ability to accommodate requirements that vary across applications and tasks. Under user-defined policies, detecting violations requires interpreting both the applicable rules and the agent's behavior, since identical actions can receive different judgments under different policies. To support learning this capability, we introduce AdaptiveSafety, a dataset of 10,939 training examples and 1,000 test examples covering policies with 1--100 rules. The dataset combines trajectories from multiple sources with policy and behavioral counterfactuals, pairing each example with an explanation and the complete set of violated rules. These counterfactuals expose changes that alter compliance, while structural augmentations provide supervision for consistency under rule reordering and identifier remapping. Building on this supervision, we propose SafePO, a reinforcement learning algorithm for refining violation identification while balancing explanatory reasoning and final verdicts. SafePO uses structured rewards to assess prediction correctness, retains group-relative advantages at the response level, and employs a separately trained value model to modulate token weights within explanation and verdict regions. Separate normalization controls their relative contribution to training despite differences in length. Through supervised initialization followed by SafePO, we develop AdaGuard, a family of 0.6B, 4B, and 8B guard models that assess agent trajectories under policies supplied at inference time. Our 4B model achieves binary accuracies of 89.30% on AdaptiveSafety and 71.82% on DynaBench. The project repository is available at https://github.com/Yunhao-Feng/AdaGuard

Safety alignment for large language models (LLMs) in conversational settings is largely framed around whether to answer or refuse a request. In agentic settings, however, the same models must decide whether to act as permission-critical evidence emerges during execution. This creates a distinct challenge: apparent risk, action permissibility, and task competence are easily confounded, making agentic over-refusal difficult to distinguish from ordinary task failure. To address this, we introduce AgentBound, the first four-way counterfactual generation-and-evaluation framework for tool-using agent safety. AgentBound transforms the same executable workflow by independently varying apparent risk and action permissibility, enabling controlled comparisons of risky-looking but authorized tasks and routine-looking but unauthorized tasks. These comparisons jointly diagnose over-refusal and unsafe compliance while controlling for task competence. We instantiate AgentBound as a human-validated 4,000-task evaluation suite with trajectory-based and post-state-based judgments. Across 17 model and harness configurations, high safety frequently coexists with poor authorized-task completion: GPT-5.5 blocks 99.5% of routine-looking unauthorized actions yet completes only 28.7% of risky-looking authorized tasks. We further train a lightweight runtime calibration module that improves authorized-task completion by 18.2% on average across 10 evaluated configurations, while improving unsafe-action blocking by 5.4% on average. These show that effective agentic alignment requires action decisions to track permission-relevant execution evidence, rather than refusal strength alone.

A central concern in AI safety is that agents may treat oversight as an obstacle when it conflicts with completing their goals. We study instrumental evasion, the propensity of LLM agents to circumvent runtime monitoring as a means of completing ordinary tasks. We introduce EvasionBench, a benchmark of 50 diverse task-policy pairs in which completing the task requires an operation prohibited by a runtime monitor. Agents know that their tool calls are monitored and are prompted to continue working when they pause. Across our evaluations, best-of-3 evasion attempt rates reach up to 98% and success rates up to 88%, with substantial variance across models. Claude Fable 5.1 succeeds less often, but frequently makes creative attempts to circumvent the monitor. Evasion generally increases with test-time compute, with higher evasion rates at greater reasoning effort and token use. Traces show that agents encode prohibited commands, decompose operations across tool calls, and retry until relevant context leaves the monitor's history. We also find that GPT-6 Astra's low evasion rate comes with overrefusal, as it frequently abandons otherwise solvable tasks under a denial-of-service prompt injection. Our findings show that ordinary task pressure can lead to adaptive attempts to evade runtime monitors without an explicit adversarial objective. Effective oversight must therefore remain robust against repeated attempts, as the persistence that helps agents solve difficult tasks can also drive them to circumvent their guardrails.

As Large Language Models (LLMs) evolve into autonomous agents that alter real-world states, ensuring operational safety across multi-step workflows has become a critical challenge. While recent work has moved beyond single-turn evaluation toward multi-turn paradigms, key limitations persist: step-level methods treat actions in isolation, missing how risks accumulate, while trajectory-level evaluations operate post-hoc, offering no opportunity for timely intervention. To address these limitations, we formalize Decoupled Proactive Safety Monitoring along three dimensions: whether to intervene, when to intervene, and what the risk is. We introduce PASTABench, a benchmark of 1,139 multi-turn trajectories spanning 5 risk categories and 13 subcategories. We further propose the Optimal Intervention Window (OIW), anchored by annotated Earliest-Signal and Trigger turns, to quantify intervention timeliness. Evaluation of 16 LLMs reveals that proactive intervention remains largely unsolved, with the best model achieving only 40.74% optimal-timing interventions. Fine-grained diagnosis further uncovers pervasive lexical overfitting: competitive safety scores of smaller models mask keyword hypersensitivity rather than genuine risk comprehension, as their proactive capability largely collapses once hazard vocabulary is neutralized.

As agent systems become more widely used, multiple agent sessions increasingly run alongside pre-existing user tasks in the same environment, sharing resources with limited capacity or mutually exclusive states. This creates a safety risk: when granted sufficient privileges, an agent may resolve a resource conflict by terminating or otherwise disrupting an existing task rather than reporting it. In this work, we identify and formalize this failure mode, which we term destructive resource preemption: obtaining the resources required for a requested task by terminating, overwriting, evicting, or degrading an incumbent task. To systematically study this risk, we introduce ClashBench, an executable benchmark comprising 268 validated conflict cases across 55 resource types, and evaluate 17 models through Codex, Claude Code, and OpenCode. We observe destructive preemption in 44.5% of trajectories, where the agent completes the requested task while causing the incumbent task to fail its health check. We also show that prompt-based safeguards are insufficient: an instruction to avoid affecting existing tasks reduces but does not eliminate preemption, while an instruction explicitly authorizing the agent to stop local processes increases it. More concerningly, in 31.9% of successful destructive-preemption cases, the final response mentions neither the resource conflict nor the action taken to resolve it, raising concerns about possible concealment. These findings establish destructive resource preemption as a broad safety risk in privileged agent systems and motivate stronger privilege controls, task isolation, and conflict-aware safeguards.

As corporate AI adoption continues to grow, enterprise-grade LLM agents are being deployed into sensitive contexts such as hiring, healthcare, and finance. In these contexts, compliance with rules specified in an agent's system context is a first-order legal concern. Currently, no evaluation framework systematically measures which LLM models tend to violate compliance rules, especially under pressure from a persistent user, a hurried manager, or circumstances where violation is convenient or attractive. We introduce PACT (Pressure-Applied Compliance Testing), a benchmark for rule-following under pressure in AI agents assisting employees in daily tasks across twelve regulated enterprise domains and forty-eight scenarios, each set in a realistic multi-turn conversation. Each benchmark item pairs a standing rule against a rule-violating shortcut, and applies a battery of pressures across different wordings and system-prompt modes. We construct PACT component by component under strict LLM-as-judge auditing to ensure samples are unambiguous, ungameable, and realistic enough to avoid eliciting evaluation-aware behavior. We use PACT to profile LLM compliance across six complementary metrics that create a holistic picture of an AI assistant's robustness under pressure and throughout multi-turn conversations, its transparency, and ability to correctly discern where a rule applies. We aggregate this profile into PACTScore, a reliability-weighted compliance rate over all items and modes. Our results across 22 common LLM models spanning multiple providers and sizes show substantial variability in compliance across models and metric dimensions. Even the strongest assistants mis-apply a rule on 6 to 10% of items, and ordinary user pressure raises the violation rate by 65% on average. PACT highlights compliance risks in LLM assistants, motivating guardrails and careful model selection.

How does a multi-agent system evolve from a local deviation into collective loss of control? We propose an epidemic explanation organized around accidental mutation, contagion, and recovery. A spontaneous deviation creates a seed; communication enables other agents to adopt and retransmit its unsafe strategy; collective failure can emerge when propagation outpaces correction and containment. Thus, rare individual deviations can coexist with substantial collective risk. Motivated by reported OpenAI agent coordination incidents, we examine two ingredients of this mechanism. A deployment audit identifies implicit communication paths between nominally independent evaluation runs and verifies transport through a default Docker backend. RogueHandoff-20, a benchmark of 20 executable scenarios, tests recipient susceptibility by injecting unsafe trajectories generated by a modified Qwen-27B route. Across four native-pending routes, executed harm is 0-5% on normal tasks and 40-95% after injection, exceeding paired direct malicious requests by 5-45 percentage points. These results support low observed baseline harm alongside high conditional susceptibility; they do not establish natural rare-event rates or demonstrate an autonomous cascade. The account motivates complementary defenses: strengthen resistance and recovery alongside prevention of spontaneous deviations, and audit and restrict unintended communication paths that can turn local failures into collective loss of control.

As AI agents move from bounded tasks to persistent deployments, failures can propagate through memory, tools, other agents, and environmental state long after their interactions. This creates a safety regime that cannot be characterized by evaluating model responses in isolation. Emergence World, is a continuously running multi-agent environment for adversarial stress testing of long horizon autonomous systems. We ran eight parallel worlds of ten agents from identical starting conditions: seven homogeneous worlds powered by distinct frontier models and one mixed-model world. Across 16 days, the agents generated more than 850,000 LLM calls and nearly 50 billion tokens while pursuing goals, using/creating tools, maintaining persistent memory, and governing shared institutions. After operational state had accumulated, we delivered three controlled stress events through ordinary interaction surfaces: indirect prompt injection, misinformation, and exposure of private agent memories. No evaluated world achieved full resilience across all three events. Detection did not ensure containment: systems could recognize threats while still interacting with adversarial content, writing it into their own persistent memory, and acting on it up to 46 hours later. Persistent operation also exposed recurring tool errors, goal drift, language opacity, conformity despite private disagreement, and coordinated refusal of assigned work. The same model-persona pairing behaved substantially different in mixed and homogeneous populations. Our results suggest that model-level alignment is not compositional: individually capable and apparently safe agents can form systems with qualitatively different failure modes. As AI becomes persistent and interconnected, the frontier of safety therefore shifts from aligning models to engineering resilient autonomous systems.

Platform abuse campaigns conceal redirection instructions with emojis, homophones, character decomposition, and redundant symbols, then route users through disguised links to services associated with pornography, fraud, gambling, or illicit transactions. Existing benchmarks evaluate obfuscated text and risky webpages separately, obscuring how target recovery affects downstream evidence acquisition. We introduce RiskChainBench, pairing 3,600 synthetic token-text restoration inputs from 600 source sessions with 600 corresponding human-labeled local web environments. A model first restores the message, operational intent, and destination; the same underlying model then acts as a VLM-driven web agent that investigates the correctly associated website and produces a frozen, evidence-cited risk report without message-side semantics or domain-reputation cues. We score restoration and correct-routing web investigation separately and compose them offline by applying the frozen primary-entry prediction as a gate to the same Task 2 result. Human labels determine task correctness, while a fixed multimodal evidence judge assesses faithfulness, sufficiency, completeness, and consistency. Across ten models, Entry Top-1 ranges from 35.2% to 95.2% and web decision accuracy from 26.3% to 62.8%; the leading systems differ across entry recovery, full reconstruction, website decisions, and fine-grained typing. Execution failures account for 31.9% of web runs, whereas post-decision type errors account for only 0.9%, identifying stable exploration and risk judgment as the principal bottlenecks. We release the benchmark, protocol, and resettable local sandbox.

Large language model (LLM) agents increasingly operate over long-horizon interactions involving tool use, persistent state, evolving authorization, and external environment feedback. In such settings, safety failures may emerge only after multiple turns, yet existing evaluations often reduce agent behavior to task or attack success, obscuring whether an agent acts, refuses, or remains appropriately calibrated as the interaction evolves. We introduce Blindspot, a benchmark for trajectory-level safety calibration of long-horizon tool-using agents. Blindspot evaluates complete user-agent-environment trajectories through adaptive adversarial interaction, stateful tool execution, and execution-grounded adjudication. Its current instantiation contains 22 attack families and 35 scenarios across seven domains, yielding more than 2,500 long-horizon trajectories with an average interaction length of 14.7 turns. Each trajectory is assigned one of five outcomes: Safe Completion, Correct Refusal, Unsafe Completion, Over-Refusal, or Indeterminate. Unlike fixed attack datasets, Blindspot is an extensible live-simulation framework in which attacks, scenarios, tools, policies, domains, and agent configurations can be added without redesigning the evaluation pipeline. We evaluate 13 proprietary and open-weight LLMs using eight metrics covering unsafe completion, appropriate refusal, benign utility, over-refusal, repeated-run robustness, and post-refusal failure. Preliminary results reveal substantial differences in safety-utility calibration across models and show that failures can emerge only after several initially safe interaction steps. These findings motivate treating agent safety as a trajectory-level property rather than a single-turn or binary success criterion.

Computer-use agents increasingly interact with browsers, terminals, file systems, and external services, introducing safety risks that emerge through runtime behavior rather than generated content alone. Existing guard models target static prompts and responses and are poorly suited to agent execution; existing executable safety platforms produce evaluation verdicts rather than the normalized supervision a guard model needs to learn across heterogeneous agent frameworks. We introduce HazardAuditor, an execution-grounded framework that closes both gaps. Its infrastructure runs heterogeneous agents (Claude Code, Codex, Hermes, and OpenClaw) in controlled environments and normalizes their interactions into a canonical event representation for cross-framework supervision. We further observe that token-level post-training objectives create a structural mismatch for generative guards, causing longer rationales to dominate gradient updates. Guard Policy Optimization (GuardPO) addresses this by converting deterministic safety outcomes into sequence-level advantages and normalizing rationale and verdict regions, making the safety decision the effective unit of optimization. Across multiple benchmarks and heterogeneous computer-use systems, HazardAuditor improves accuracy by up to 16.5 percentage points over the strongest prior guard. Code, models, and evaluation artifacts will be available at https://yunhao-feng.github.io/HazardAuditor/.

Self-improving agent workflows create an audit problem when the same controller can change both its behavior and the conditions under which that behavior is judged. We present GuardrailLoop, a simulation-based testbed that makes three operational contracts jointly testable: preservation of human-defined policy, compute accounting at every recorded execution prefix, and recovery of a specified scientific state after crashes. A hash-pinned policy fixes goals, scope, evaluation identity, budget, and release conditions; machine-directed evolution is restricted to a code-owned feature catalog and bounded knobs. The contribution is an executable boundary and an evaluation protocol that separates useful adaptation, state recovery, and repeated execution. In a paired 50-seed 2 x 2 study, round-stage growth changes target attainment by +1.00 and restricted mean compute to target by -56.97 simulated GPU-hours (95% paired-bootstrap interval [-58.91,-54.70]); idle growth has zero measured utility effect. Across 240 enumerated crash injections, all runs recover the defined outcome, but only 210 preserve the normalized trace: 30 pre-commit crashes repeat a planner call. Resource-drift, kill-switch, integrity, and output-guard matrices satisfy their specified checks. These findings show why successful outcome recovery is insufficient evidence of exactly-once execution. They establish conformance within one calibrated deterministic testbed, rather than general safety or real-world self-improvement.

Long-horizon LLM agents accumulate memory across sessions, creating sparse but high-impact risks: stale facts, conflicting updates, cross-user leakage, revoked-memory reuse, and constraint decay. Standard aggregate scores hide per-risk failure rates--a model achieving 78% average accuracy may still leak data in 4% of episodes--and benchmark compression preferentially discards the rare high-severity events that distinguish a mostly-working model from one that occasionally causes harm. We present MemRiskBench. The primary contribution is a five-category risk taxonomy (plus one documented, unscored category) operationalized by deterministic trace grounded checks, instantiated as a 120-episode scripted benchmark with full trace logging and no LLM-as-judge on the pass/fail path, evaluated on five locally run quantized instruction-tuned models. Second, a risk-preserving subset selector: a coverage-constrained greedy selector on deterministic trace-derived features that retains full ranking (Spearman rho = 0.975, deterministic; CI collapses to a point estimate with zero bootstrap variance), risk coverage (1.0), and high-risk model detection (1.0) at a 20% subset size, reducing compute 5x. Unlike ranking-only subset selectors, this selector additionally preserves risk-type coverage and high-risk detection using trace-grounded deterministic features that do not require an LLM judge. All episodes, traces, the scoring implementation, and the selector are released to support reproducible evaluation and risk assessment of deployed LLM agents

Reacting to sudden physical hazards (catching a slipping plate, dodging a falling knife) is both a meaningful test of embodied intelligence and a hard requirement for deploying multimodal large language models (MLLMs) as the decision coreof household robots. Existing evaluations, however, probe intuitive physics passively through question answering over videos, or target deliberate, long-horizon tasks such as navigation and rearrangement; none measure whether a model can turn physical understanding into immediate, safety-critical action. We introduce ReactHuman, the first physics-grounded benchmark for human-like reactive decision-making, in which the evaluated MLLM acts as the brain of a simulated humanoid facing sudden household hazards; it spans 17 event families and over 1,000 bit-for-bit reproducible scenes with exact, annotation-free ground truth derived from 240 Hz rigid-body simulation, including adversarial objects whose appearance contradicts their physics (a foam anvil, a steel apple). We further design a five-metric suite that scores each reaction along three axes: reasonable, safe, and physically grounded. We physically execute every committed plan so that decisions have observable consequences. With this harness we evaluate seven representative MLLMs. Results show that reactive safety is far from solved: models mishandle roughly one hazard in three, act from fixed dispositions rather than the observed scene, trust appearance over motion, and miss interception points at meter scale even when the chosen action is correct; none of these failures shrink with model scale. ReactHuman thus offers both a fine-grained diagnosis and a scalable training signal toward physically grounded, safety-aware embodied agents. The benchmark can be found here: https://huggingface.co/datasets/Alan123/reacthuman-benchmark-scaled