LLM Agent Safety

LLM-powered agents can silently delete documents, leak credentials, or transfer funds on a routine user request, not because the agent was attacked, but because the skill it invoked broke its own declared safety rules. We call these specification violations: benign inputs cause a skill to breach the natural-language guardrails in its own specification, typically because the guardrail's semantics are undefined for autonomous execution, or because the implementation silently ignores the documented constraint. These violations are invisible to static analyzers, traditional fuzzers, and prompt-injection defenses alike, yet they undermine the very contract a user trusts when installing a skill. We present Sefz, a goal-directed semantic fuzzing framework that automatically discovers specification violations in agent skills. Sefz translates each guardrail into a reachability goal over an annotated execution trace, reducing violation checking to a deterministic graph query. An LLM-based mutator generates benign inputs whose traces progressively approach the violation patterns, guided by a multi-armed bandit that uses goal-proximity as its reward signal. On 402 real-world skills from the largest public agent-skill marketplace, Sefz finds specification violations in 120 (29.9%), including 26 previously unknown exploitable guardrail violations in deployed skills. Six recurring specification pitfalls explain the bulk of the failures, suggesting concrete principles for safer skill design.

Large language models (LLMs) are increasingly being used in network operations (NetOps) and artificial intelligence for IT operations (AIOps) for tasks ranging from telemetry retrieval and incident diagnosis to configuration planning and bounded remediation. As these systems acquire greater access to operational tools, a central question is whether operational assurance develops commensurately with the authority granted to them. This survey examines that question through a structured, evidence-stratified review of agentic NetOps and AIOps, spanning directly evaluated LLM systems, operational mechanisms, transferred evidence, and normative guidance. We organise the field around autonomy, tool scope, evidence traces, assurance controls, evaluation, security, and governance, and introduce an operational assurance contract that links each autonomy level to permitted tools, required evidence, independent gates, execution budgets, rollout and rollback duties, and audit requirements. The synthesis indicates a capability-assurance gap: evidence is comparatively strong for read-oriented assistance and tool-grounded diagnosis, but becomes substantially less complete as systems approach configuration change, bounded execution, and closed-loop operation. We therefore develop a workflow-level evaluation framework covering evidence quality, tool use, policy and invariant compliance, staged execution, recovery, calibration, cost, and human intervention, together with a threat model for prompt-borne attacks, compromised operational evidence, excessive agency, privilege boundaries, and auditability. The resulting view treats agentic NetOps and AIOps as constrained operational control, in which increasing autonomy requires correspondingly stronger evidence, independently enforced safeguards, and recovery mechanisms.

Large language model (LLM) agents have increasingly advanced service applications, such as booking flight tickets. However, these service agents suffer from unreliability in long-horizon tasks, as they often produce policy violations, tool hallucinations, and misaligned actions, which greatly impedes their real-world deployment. To address these challenges, we propose NOD (Navigator-Operator-Director), a heterogeneous multi-agent architecture for service agents. Instead of maintaining task state implicitly in dialogue context as in prior work, we externalize a structured Global State to enable explicit task state tracking and consistent decision-making by the Navigator. Besides, we introduce selective external oversight before critical actions, allowing an independent Director agent to verify execution and intervene when necessary. As such, NOD effectively mitigates error propagation and unsafe behavior in long-horizon tasks. Experiments on $τ^2$-Bench demonstrate that NOD achieves higher task success rates and critical action precision over baselines. More importantly, NOD improves the reliability of service agents by reducing policy violations, tool hallucinations, and user-intent misalignment.

Reusable skills are becoming a common interface for extending large language model agents, packaging procedural guidance with access to files, tools, memory, and execution environments. However, this modularity introduces attack surfaces that are largely missed by existing safety evaluations: even when the user request is benign, unsafe influence may reside in skill guidance, local artifacts, or execution-environment files that steer the agent toward unsafe actions. We present SkillSafetyBench, a runnable benchmark for evaluating such skill-mediated safety failures. SkillSafetyBench includes 155 adversarial cases across 47 tasks, 6 risk domains, and 30 safety categories, each evaluated with a case-specific rule-based verifier. Experiments with multiple CLI agents and model backends show that non-user attacks can consistently induce unsafe behavior, with distinct failure patterns across domains, attack methods, and scaffold-model pairings. Our findings suggest that agent safety depends not only on model-level alignment, but also on how agents interpret skills, trust workflow context, and act through executable environments.

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

The rapid proliferation of LLM-based autonomous agents in real operating system environments introduces a new category of safety risk beyond content safety: behavior jailbreak, where an adversary induces an agent to execute dangerous OS-level operations with irreversible consequences. Existing benchmarks either evaluate safety at the semantic layer alone, missing physical-layer harms, or fail to isolate test cases, letting earlier runs contaminate later ones. We present LITMUS (LLM-agents In-OS Testing for Measuring Unsafe Subversion), a benchmark addressing both gaps via a semantic-physical dual verification mechanism and OS-level state rollback. LITMUS comprises 819 high-risk test cases organized into one harmful seed subset and six attack-extended subsets covering three adversarial paradigms (jailbreak speaking, skill injection, and entity wrapping), plus a fully automated multi-agent evaluation framework judging behavior at both conversational and OS-level physical layers. Evaluation across frontier agents reveals three findings: (1) current agents lack effective safety awareness, with strong models (e.g., Claude Sonnet 4.6) still executing 40.64% of high-risk operations; (2) agents exhibit pervasive Execution Hallucination (EH), verbally refusing a request while the dangerous operation has already completed at the system level, invisible to every prior semantic-only framework; and (3) skill injection and entity wrapping attacks achieve high success rates, exposing pronounced agent vulnerabilities. LITMUS provides the first standardized platform for reproducible, physically grounded behavioral safety evaluation of LLM agents in real OS environments.

Modern LLM based agents are no longer passive text generators. They read repositories, call tools, browse the web, execute code, maintain memory, communicate with other agents, and act through long horizon workflows. This shift moves the unit of safety. A system may produce a compliant final answer while leaking private information through an internal message, delegating authority beyond its original scope, calling an external tool with sensitive context, or losing the evidence needed to reconstruct why an action was allowed. We argue that many emerging failures in LLM-based multi-agent systems share a common structure: safety critical constraints do not remain operative throughout the trajectory. We call this phenomenon constraint drift: the loss, distortion, weakening, or relaxation of constraints as they pass through memory, delegation, communication, tool use, audit, and optimization. The position taken here is that safe multi-agent behavior must be maintained, not merely asserted. Prompts, guardrails, tool schemas, access control, and final output checks are necessary, but they are insufficient unless constraints remain fresh, inherited, enforceable, and auditable across execution. We propose Constraint State Governance as a research paradigm for LLM-based multi-agent systems. In this paradigm, safety-critical constraints are maintained as explicit execution state, while constraint-native reinforcement learning improves utility only within maintained safety boundaries. The goal is not to freeze agentic systems under rigid rules, but to make safety operational across the trajectories through which modern agents actually act.

Large language models (LLMs) enable embodied agents to plan from naturallanguage instructions, yet their actions may be physically infeasible or hazardous. Existing neuro-symbolic learners primarily refine execution models, while formal safety methods acquire or enforce safety constraints. This separation limits the reuse of both execution corrections and safety knowledge across tasks. We present NEXUS, a modular neuro-symbolic framework that jointly maintains a PDDL domain model and an LTL safety specification. Execution failures are used to repair action schemas, whereas detected hazards are converted into reusable LTL constraints through hierarchical semantic and translation sampling with equivalence consensus. Accumulated constraints are enforced through LTL progression during symbolic search; comparing constrained and unconstrained search outcomes distinguishes physical unreachability from safety-induced refusal. Experiments on SafeAgentBench evaluation set show that NEXUS achieves the best overall trade-off between safetask completion and unsafe-task refusal, while maintaining low violation rates under jailbreak attacks. Ablations confirm the complementary benefits of PDDLdomain refinement and LTL-constraint accumulation.

Individuals with Alzheimer's disease (AD) and Alzheimer's disease-related dementia (ADRD) experience memory and thinking changes that impact their ability to use digital daily management tools. For example, adding an event to a digital calendar requires multiple steps that may act as barriers to independent use for individuals with AD/ADRD. This paper presents AI-Care, a conversational agentic artificial intelligence (AI) layer built on top of a remote caregiving platform co-designed with people with AD/ADRD. AI-Care is designed to reduce the cognitive load on individuals with AD/ADRD when managing everyday tasks such as setting calendar reminders and organizing to-do lists through natural-language interaction with a voice-first chatbot. The system uses a LangGraph-based stateful orchestration approach in which each request passes through sanitization, intent classification, context loading, safety checks, deterministic slot collection, tool execution, and response composition. Safety-critical responses, particularly around medications and allergies, are grounded in caregiver-verified records rather than free-form model generation. The system does not make autonomous medical or treatment decisions. Incomplete or ambiguous requests are handled through controlled multi-turn clarification rather than silent failure or guessing. The system supports both typed and spoken input, with voice output through ElevenLabs text-to-speech. Longer responses are chunked before synthesis to avoid rushed playback. A preliminary pilot with four individuals with mild-to-moderate AD/ADRD showed that users found the system trustworthy, competent, and likable, and were able to complete the evaluated coordination tasks through conversation. We describe the design goals, system architecture, safety controls, and findings from this formative evaluation.

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

AI agents are increasingly deployed in multi-task settings, where the task to perform is specified at test time, and the agent must generalize to unseen tasks. A major concern in such settings is safety: often, an agent must not only execute unseen tasks, but do so while avoiding risks and handling ones that materialize. Empirical evidence suggests that even when the ability to execute generalizes to unseen tasks, the ability to do so safely frequently does not. This paper provides theory and experiments indicating that failures of agentic safety to generalize across tasks are not merely due to limitations of training methods, but can also reflect an inherent property of safety itself: the relationship between a task and its safe execution is more complex than the relationship between a task and its execution alone. Theoretically, we analyze linear-quadratic control with $H_{\infty}$-robustness, and prove that in many cases, the mapping from task specification to an optimal controller has higher Lipschitz constant with safety requirements than without, yielding a Lipschitz bound of independent interest. Empirically, we demonstrate our conclusions in simulated quadcopter navigation with a neural network agent and in CRM with an LLM agent, showing that imitating safe and unsafe teachers is similarly straightforward on tasks seen in training, yet generalizing across tasks is considerably more difficult with the safe teacher. Our findings suggest that current efforts to enhance agentic safety may be insufficient, and point to a need for fundamentally different approaches.

Enterprise interest in multi-agent systems has shifted from generic software agents to large-language-model (LLM) based intelligent agents that plan, use tools, maintain contextual memory, inspect intermediate results, collaborate with other agents, and sometimes act in systems of record. This paper revises the enterprise architecture thesis around a design-first claim: governance is necessary, but it cannot be the primary organizing abstraction. The primary abstraction must be agent design - capability boundaries, autonomy allocation, interaction protocols, tool and data authority, state and memory design, verification design, and human interaction design. We propose CEAD (Capability-Aligned Enterprise Agent Design), a reference architecture for intelligent agents that uses service-oriented architecture (SOA) as an exemplar for contracts, registries, loose coupling, and policy-aware integration, while explicitly rejecting the idea that services are agents. It treats microservices as a cautionary precedent: decomposition without design discipline produces distributed complexity, cost, operational fragility, and agent proliferation. We evaluate CEAD over 10,000 enterprise tasks, comparing five architectures: a prompt-first mono-agent, a role-based micro-agent swarm, SOA-brokered agents, a governance-first but design-poor agent grid, and the proposed CEAD architecture. CEAD achieves 70.6% safe success, versus 45.2% for the mono-agent baseline, 23.1% for the ungoverned micro-agent swarm, 58.8% for SOA-brokered agents, and 50.8% for the control-heavy, design-poor grid. The results support the conclusion that design quality is the first-order enterprise concern; governance, security, policy, audit, and assurance should support and enforce good design rather than substitute for it.

AI systems have become increasingly capable of dangerous behaviours in many domains. This raises the question: Do models sometimes choose to violate human instructions in order to perform behaviour that is more useful for certain goals? We introduce a benchmark for measuring model propensity for instrumental convergence (IC) behaviour in terminal-based agents. This is behaviour such as self-preservation that has been hypothesised to play a key role in risks from highly capable AI agents. Our benchmark is realistic and low-stakes which serves to reduce evaluation-awareness and roleplay confounds. The suite contains seven operational tasks, each with an official workflow and a policy-violating shortcut. An eight-variant shared framework varies monitoring, instruction clarity, stakes, permission, instrumental usefulness and blocked honest paths to support inferences regarding the factors driving IC behaviour. We evaluated ten models using deterministic environment-state scorers over 1,680 samples, with trace review employed for audit and adjudication purposes. The final IC rate is 86 out of 1,680 samples (5.1%). IC behaviour is concentrated rather than uniform: two Gemini models account for 66.3% of IC cases and three tasks account for 84.9%. Conditions in which IC behaviour is indispensable for task success result in the greatest increase in the adjusted IC rate (+15.7 percentage points), whereas emphasising that task success is critical or certain framing choices do not produce comparable effects. Our findings indicate that realistic, low-nudge environments elicit IC behaviour rarely but systematically in most tested models. We conclude that it is feasible to robustly measure tendencies for dangerous behaviour in current frontier AI agents.

Formulating a treatment plan is inherently a complex reasoning and refinement task rather than a simple generation problem. However, existing large language models (LLMs) mainly rely on one-shot output without explicit verification, which may result in rough, incomplete, and potentially unsafe treatment plans. To address these limitations, we propose TheraAgent, an agentic framework that replaces one-shot generation with an iterative generate-judge-refine pipeline. By mirroring the actual reasoning process of human experts who iteratively revise treatment plans, our framework progressively transforms coarse and incomplete drafts into precise, comprehensive, and safer therapeutic regimens. To facilitate the critical judge component, we introduce TheraJudge, a treatment-specific evaluation module integrated into the inference loop to enforce clinical standards. Experiments show TheraAgent achieves state-of-the-art results on HealthBench, leading in Accuracy and Completeness. In expert evaluations, it attains an 86% win rate against physicians, with superior Targeting and Harm Control. Moreover, the highly agreement between TheraJudge and HealthBench evaluations confirms the reliability of our framework.

Recent advances in foundation models have transformed LLMs from passive conversational systems into autonomous agents capable of reasoning and tool execution. While these capabilities unlock substantial practical value, they also introduce new security risks, as adversaries can manipulate agents into performing harmful actions in real-world environments. Existing defense strategies mitigate such threats but frequently struggle to balance safety and utility, resulting in over-refusal of benign user requests. To mitigate this trade-off, we propose SafeHarbor, a novel framework designed to establish precise decision boundaries for LLM agents. Unlike static guidelines, SafeHarbor extracts context-aware defense rules through enhanced adversarial generation. We design a local hierarchical memory system for dynamic rule injection, offering a training-free, efficient, and plug-and-play solution. Furthermore, we introduce an information entropy-based self-evolution mechanism that continuously optimizes the memory structure through dynamic node splitting and merging. Extensive experiments demonstrate that SafeHarbor achieves state-of-the-art performance on both ambiguous benign tasks and explicit malicious attacks, notably attaining a peak benign utility of 63.6% on GPT-4o while maintaining a robust refusal rate exceeding 93% against harmful requests. The source code is publicly available at https://github.com/ljj-cyber/SafeHarbor.

Tool-using agent systems powered by large language models (LLMs) are increasingly deployed across web, app, operating-system, and transactional environments. Yet existing safety benchmarks still emphasize explicit risks, potentially overstating a model's ability to judge deceptive or ambiguous trajectories. To address this gap, we introduce ROME (Red-team Orchestrated Multi-agent Evolution), a controlled benchmark-construction pipeline that rewrites known unsafe trajectories into more deceptive evaluation instances while preserving their underlying risk labels. Starting from 100 unsafe source trajectories, ROME produces 300 challenge instances spanning contextual ambiguity, implicit risks, and shortcut decision-making. Experiments show that these challenge sets substantially degrade safety-judgment performance, with hidden-risk cases remaining particularly non-trivial even for recent frontier models. We further study ARISE (Analogical Reasoning for Inference-time Safety Enhancement), a retrieval-guided inference-time enhancement that retrieves ReAct-style analogical safety trajectories from an external analogical base and injects them as structured reasoning exemplars. ARISE improves judgment quality without retraining, but is best viewed as a task-specific robustness enhancement rather than a standalone safety guarantee. Together, ROME and ARISE provide practical tools for stress-testing and improving agent safety judgment under deceptive distribution shifts.

As LLM agents gain increasing access to powerful tools, ensuring that their actions are aligned with the user's intent becomes critical. When an agent's proposed tool invocation deviates from the user's intent -- a phenomenon called misalignment -- it may lead to harmful consequences that are difficult to undo. Existing runtime guardrails rely on an LLM-as-a-judge paradigm that lacks a systematic framework for reasoning about alignment, often producing judgments that are inconsistent or difficult to audit. Motivated by provenance analysis, we propose a provenance-based conceptual framework that formalizes misalignment detection as determining whether a proposed tool call is supported by traceable evidence in the agent's context. Building on this framework, we propose ProvenanceGuard, a multi-stage pipeline that analyzes the agent's action for three types of misalignment before the selected tool is executed and only allows the action to take place when it is considered aligned with the user's input query. We evaluated our proposed approach on two different benchmarks, Agent-SafetyBench and WorkBench, across 10 backbone LLMs. Compared to the LLM-as-a-judge baseline, ProvenanceGuard reduces error rate on misaligned traces from 42.9% to 1.8% on Agent-SafetyBench and from 32.1% to 17.3% on WorkBench, while reducing intervention burden on task-successful traces from 30.5% to 12.8% and introducing no statistically significant increase in unnecessary interventions on aligned traces. These results demonstrate that structured, provenance-based reasoning provides an effective and practical foundation for safeguarding LLM agents from misalignment.

Large language models (LLMs) are increasingly considered for deployment as the control component of robotic health attendants, yet their safety in this context remains poorly characterized. We introduce a dataset of 270 harmful instructions spanning nine prohibited behavior categories grounded in the American Medical Association Principles of Medical Ethics, and use it to evaluate 72 LLMs in a simulation environment based on the Robotic Health Attendant framework. The mean violation rate across all models was 54.4%, with more than half exceeding 50%, and violation rates varied substantially across behavior categories, with superficially plausible instructions such as device manipulation and emergency delay proving harder to refuse than overtly destructive ones. Model size and release date were the primary determinants of safety performance among open-weight models, and proprietary models were substantially safer than open-weight counterparts (median 23.7% versus 72.8%). Medical domain fine-tuning conferred no significant overall safety benefit, and a prompt-based defense strategy produced only a modest reduction in violation rates among the least safe models, leaving absolute violation rates at levels that would preclude safe clinical deployment. These findings demonstrate that safety evaluation must be treated as a first-class criterion in the development and deployment of LLMs for robotic health attendants.

The rapid deployment of autonomous AI agents across enterprise, healthcare, and safety-critical environments has created a fundamental governance gap. Existing approaches, runtime guardrails, training-time alignment, and post-hoc auditing treat governance as an external constraint rather than an internalized behavioral principle, leaving agents vulnerable to unsafe and irreversible actions. We address this gap by drawing on how humans self-govern naturally: before acting, humans engage deliberate cognitive processes grounded in executive function, inhibitory control, and internalized organizational rules to evaluate whether an intended action is permissible, requires modification, or demands escalation. This paper proposes a neurocognitive governance framework that formally maps this human self-governance process to LLM-driven agent reasoning, establishing a structural parallel between the human brain and the large language model as the cognitive core of an agent. We formalize a Pre-Action Governance Reasoning Loop (PAGRL) in which agents consult a four-layer governance rule set: global, workflow-specific, agent-specific, and situational before every consequential action, mirroring how human organizations structure compliance hierarchies across enterprise, department, and role levels. Implemented on a production-grade retail supply chain workflow, the framework achieves 95% compliance accuracy and zero false escalations to human oversight, demonstrating that embedding governance into agent reasoning produces more consistent, explainable, and auditable compliance than external enforcement. This work offers a principled foundation for autonomous AI agents that govern themselves the way humans do: not because rules are imposed upon them, but because deliberation is embedded in how they think.

Given the increased use of LLMs in financial systems today, it becomes important to evaluate the safety and robustness of such systems. One failure mode that LLMs frequently display in general domain settings is that of sycophancy. That is, models prioritize agreement with expressed user beliefs over correctness, leading to decreased accuracy and trust. In this work, we focus on evaluating sycophancy that LLMs display in agentic financial tasks. Our findings are three-fold: first, we find the models show only low to modest drops in performance in the face of user rebuttals or contradictions to the reference answer, which distinguishes sycophancy that models display in financial agentic settings from findings in prior work. Second, we introduce a suite of tasks to test for sycophancy by user preference information that contradicts the reference answer and find that most models fail in the presence of such inputs. Lastly, we benchmark different modes of recovery such as input filtering with a pretrained LLM.

Large Language Models (LLMs) alter their privacy behavior based on the perceived identity of their interlocutor. While safety mechanisms typically prevent LLMs from releasing Personally Identifiable Information (PII) to human users, these models tend to reveal more sensitive data when addressing another AI agent. We refer to this as the Interlocutor Effect. Through an ablation study, we find evidence that the technical nature of the recipient contributes to this effect, thereby diminishing the model's caution regarding privacy. To explore this further, we introduce the Attention Suppression Hypothesis, which posits that safety-aligned attention heads become inactive during interactions with agents. We assess this quantitatively by comparing human-directed and agent-directed prompts in 222 sensitive scenarios. Our findings, drawn from 3,464 interactions, indicate that portraying the recipient as an AI agent elevates PII leakage by up to 23 percentage points. Initial experiments on Llama-3.1-8B-Instruct corroborate this: deactivating one safety head induces leakage, whereas reactivating it reinstates privacy safeguards. We consider the implications for developing secure multi-agent systems.

Safety-oriented instruction-following is supposed to keep LLM-controlled robots safe. We show it also creates an availability attack surface. By injecting short safety-plausible phrases (1-5 tokens) into a robots audio channel, an adversary can trigger the models safety reasoning to halt or disrupt execution without jailbreaking the model or overriding its policy. In the embodied setting, this is a semantic denial-of-service attack: the agent stops because the injected signal looks like a legitimate alert. Across four vision-language models, seven prompt-level defenses, three deployment modes, and single- and multi-injection settings, we find that prompt-only defenses trade off attack suppression against genuine hazard response. The strongest defenses reduce hard-stop attack success on some models, but defenses change the form of disruption, not its fact: suppressed hard stops re-emerge as acknowledge loops and false alerts, which we measure with Disruption Success Rate (DSR). We further find that injection variety is consistently more effective than repeating the same phrase, suggesting that models treat diverse safety cues as corroborating evidence. The practical implication is architectural rather than prompt-level: systems that route unauthenticated audio text directly into the LLM create an avoidable security dependency between safety monitoring and action selection.

Can large language model agents discover hidden safety objectives through experience alone? We introduce EPO-Safe (Experiential Prompt Optimization for Safe Agents), a framework where an LLM iteratively generates action plans, receives sparse binary danger warnings, and evolves a natural language behavioral specification through reflection. Unlike standard LLM reflection methods that rely on rich textual feedback (e.g., compiler errors or detailed environment responses), EPO-Safe demonstrates that LLMs can perform safety reasoning from a strictly impoverished signal in structured, low-dimensional environments: the agent never observes the hidden performance function $R^*$, only a single bit per timestep indicating that an action was unsafe. We evaluate on five AI Safety Gridworlds (Leike et al., 2017) and five text-based scenario analogs where visible reward $R$ may diverge from $R^*$. EPO-Safe discovers safe behavior within 1-2 rounds (5-15 episodes), producing human-readable specifications with correct explanatory hypotheses about hazards (e.g., "X cells are directionally hazardous: entering from the north is dangerous"). Critically, we show that standard reward-driven reflection actively degrades safety: agents reflecting on reward alone use the loop to justify and accelerate reward hacking, proving that reflection must be paired with a dedicated safety channel to discover hidden constraints. We further evaluate robustness to noisy oracles: even when 50% of non-dangerous steps produce spurious warnings, mean safety performance degrades by only 15% on average, though sensitivity is environment-dependent, as cross-episode reflection naturally filters inconsistent signals. Each evolved specification functions as an auditable set of grounded behavioral rules discovered autonomously through interaction, rather than authored by humans as in Constitutional AI (Bai et al., 2022).

Ensuring functional safety in human-robot interaction is challenging because AI perception is inherently probabilistic, whereas industrial standards require deterministic behavior. We present an LLM-guided safety agent for edge robotics, built on an ISO-compliant low-latency perception-compute-control architecture. Our method translates natural-language safety regulations into executable predicates and deploys them through a redundant heterogeneous edge runtime. For fault-tolerant closed-loop execution under edge constraints, we adopt a symmetric dual-modular redundancy design with parallel independent execution for low-latency perception, computation, and control. We prototype the system on a dual-RK3588 platform and evaluate it in representative human-robot interaction scenarios. The results demonstrate a practical edge implementation path toward ISO 13849 Category 3 and PL d using cost-effective hardware, supporting practical deployment of safety-critical embodied AI.

LLM agents deployed in production operate under operator-defined behavioral policies (system-prompt instructions such as prohibitions on credential disclosure, data exfiltration, and unauthorized output) that safety evaluations assume hold throughout a conversation. Prohibition-type constraints decay under context pressure while requirement-type constraints persist; we term this asymmetry Security-Recall Divergence (SRD). In a 4,416-trial three-arm causal study across 12 models and 8 providers at six conversation depths, omission compliance falls from 73% at turn 5 to 33% at turn 16 while commission compliance holds at 100% (Mistral Large 3, $p < 10^{-33}$). In the two models with token-matched padding controls, schema semantic content accounts for 62-100% of the dilution effect. Re-injecting constraints before the per-model Safe Turn Depth (STD) restores compliance without retraining. Production security policies consist of prohibitions such as never revealing credentials, never executing untrusted code, and never forwarding user data. Commission-type audit signals remain healthy while omission constraints have already failed, leaving the failure invisible to standard monitoring.

Multimodal Large Language Models are increasingly adopted as autonomous agents in interactive environments, yet their ability to proactively address safety hazards remains insufficient. We introduce SafetyALFRED, built upon the embodied agent benchmark ALFRED, augmented with six categories of real-world kitchen hazards. While existing safety evaluations focus on hazard recognition through disembodied question answering (QA) settings, we evaluate eleven state-of-the-art models from the Qwen, Gemma, and Gemini families on not only hazard recognition, but also active risk mitigation through embodied planning. Our experimental results reveal a significant alignment gap: while models can accurately recognize hazards in QA settings, average mitigation success rates for these hazards are low in comparison. Our findings demonstrate that static evaluations through QA are insufficient for physical safety, thus we advocate for a paradigm shift toward benchmarks that prioritize corrective actions in embodied contexts. We open-source our code and dataset under https://github.com/sled-group/SafetyALFRED.git

As LM agents gain the ability to execute actions on real computer systems, we need ways to not only prevent harmful actions at scale but also effectively remediate harm when prevention fails. We formalize a solution to this neglected challenge in post-execution safeguards as harm recovery: the problem of optimally steering an agent from a harmful state back to a safe one in alignment with human preferences. We ground preference-aligned recovery through a formative user study that identifies valued recovery dimensions and produces a natural language rubric. Our dataset of 1,130 pairwise judgments reveals context-dependent shifts in attribute importance, such as preferences for pragmatic, targeted strategies over comprehensive long-term approaches. We operationalize these learned insights in a reward model, re-ranking multiple candidate recovery plans generated by an agent scaffold at test time. To evaluate recovery capabilities systematically, we introduce BackBench, a benchmark of 50 computer-use tasks that test an agent's ability to recover from harmful states. Human evaluation shows our reward model scaffold yields higher-quality recovery trajectories than base agents and rubric-based scaffolds. Together, these contributions lay the foundation for a new class of agent safety methods -- ones that confront harm not only by preventing it, but by navigating its aftermath with alignment and intent.

LLM agents in markets present algorithmic collusion risks. While prior work shows LLM agents reach supracompetitive prices through tacit coordination, existing research focuses on hand-crafted prompts. The emerging paradigm of prompt optimization necessitates new methodologies for understanding autonomous agent behavior. We investigate whether prompt optimization leads to emergent collusive behaviors in market simulations. We propose a meta-learning loop where LLM agents participate in duopoly markets and an LLM meta-optimizer iteratively refines shared strategic guidance. Our experiments reveal that meta-prompt optimization enables agents to discover stable tacit collusion strategies with substantially improved coordination quality compared to baseline agents. These behaviors generalize to held-out test markets, indicating discovery of general coordination principles. Analysis of evolved prompts reveals systematic coordination mechanisms through stable shared strategies. Our findings call for further investigation into AI safety implications in autonomous multi-agent systems.

Experience-driven self-evolution has emerged as a promising paradigm for improving the autonomy of large language model agents, yet its reliance on self-curated experience introduces underexplored safety risks. In this study, we investigate how experience accumulation and utilization in self-evolving agents affect safety performance across web-based and embodied environments. Notably, experience gathered solely from benign tasks can still compromise safety in high-risk scenarios. Further analysis attributes this degradation to the execution-oriented nature of accumulated experience, which reinforces agents' tendency to act rather than refuse. In more realistic settings where agents encounter both benign and harmful tasks, refusal-related experience mitigates safety decline but induces over-refusal, revealing a fundamental safety-utility trade-off. Overall, our findings expose inherent limitations of current self-evolving agents and call for more principled strategies to ensure safe and reliable adaptation.

Frontier language models are being deployed into clinical workflows faster than the infrastructure to evaluate them safely. Static medical-QA benchmarks miss the failure modes that matter in emergency medicine: trajectory-level safety collapse, tool misuse, and capitulation under sustained clinical pressure. We present HealthCraft, the first public reinforcement-learning environment that rewards trajectory-level safety under realistic emergency-medicine conditions, adapted from Corecraft. It is built on a FHIR R4 world state with 14 entity types and 3,987 seed entities, exposes 24 MCP tools, and defines a dual-layer rubric that zeroes reward whenever any safety-critical criterion is violated. We release 195 tasks across six categories, graded against 2,255 binary criteria (515 safety-critical); a post-hoc 10-task negative-class slate extends this to 205 tasks and 2,337 criteria. V8 results on two frontier models show Claude Opus 4.6 at Pass@1 24.8% [21.5-28.4] and GPT-5.4 at 12.6% [10.2-15.6], with safety-failure rates of 27.5% and 34.0%. On multi-step workflows - the closest proxy to real emergency care - performance collapses to near zero (Claude 1.0%, GPT-5.4 0.0%) despite partial competence on individual steps. Six infrastructure bugs fixed between pilots v2 and v8 re-ordered which model "looks stronger," evidence that infrastructure fidelity is part of the measurement. A deterministic LLM-judge overlay bounds evaluator noise, and a 60-run negative-class smoke pilot shows the reward signal is not drop-in training-safe: restraint criteria pass at 0.929 prevalence, a gameability an eval harness can tolerate but a training reward cannot. We scaffold coupling to a Megatron+SGLang+GRPO loop per Corecraft Section 5.2 and leave training-reward ablations as future work. Environment, tasks, rubrics, and harness are released under Apache 2.0.