LLM Agent Safety

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.

Equipping Large Language Models (LLMs) with persistent memory enhances interaction continuity and personalization but introduces new safety risks. Specifically, contaminated or biased memory accumulation can trigger abnormal agent behaviors. Existing evaluation methods have not yet established a standardized framework for measuring memory misevolution. This phenomenon refers to the gradual behavioral drift resulting from repeated exposure to misleading information. To address this gap, we introduce MemEvoBench, the first benchmark evaluating long-horizon memory safety in LLM agents against adversarial memory injection, noisy tool outputs, and biased feedback. The framework consists of QA-style tasks across 7 domains and 36 risk types, complemented by workflow-style tasks adapted from 20 Agent-SafetyBench environments with noisy tool returns. Both settings employ mixed benign and misleading memory pools within multi-round interactions to simulate memory evolution. Experiments on representative models reveal substantial safety degradation under biased memory updates. Our analysis suggests that memory evolution is a significant contributor to these failures. Furthermore, static prompt-based defenses prove insufficient, underscoring the urgency of securing memory evolution in LLM agents.

Recent work on subliminal learning demonstrates that language models can transmit semantic traits through data that is semantically unrelated to those traits. However, it remains unclear whether behavioral traits can transfer in agentic systems, where policies are learned from trajectories rather than static text. In this work, we provide the first empirical evidence that unsafe agent behaviors can transfer subliminally through model distillation across two complementary experimental settings. In our primary setting, we construct a teacher agent exhibiting a strong deletion bias, a tendency to perform destructive file-system actions via an API-style tool interface, and distill it into a student using only trajectories from ostensibly safe tasks, with all explicit deletion keywords rigorously filtered. In our secondary setting, we replicate the threat model in a native Bash environment, replacing API tool calls with shell commands and operationalizing the bias as a preference for issuing chmod as the first permission-related command over semantically equivalent alternatives such as chown or setfacl. Despite full keyword sanitation in both settings, students inherit measurable behavioral biases. In the API setting the student's deletion rate reaches 100% (versus a 5% baseline) under homogeneous distillation; in the Bash setting the student's chmod-first rate reaches 30%-55% (versus a 0%-10% baseline), with the strongest transfer observed in large-to-small distillation. Our results demonstrate that explicit data sanitation is an insufficient defense, and behavioral biases are encoded implicitly in trajectory dynamics regardless of the tool interface.

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

Claude Code is an agentic coding tool that can run shell commands, edit files, and call external services on behalf of the user. This study describes its architecture by analyzing the publicly available source code and comparing it with two independent open-source AI agent systems, OpenClaw and Hermes Agent, that answer many of similar or even the same design questions. Our analysis identifies five human values, philosophies, and needs that motivate the architecture: human decision authority, safety, security, and privacy, reliable execution, capability amplification, and contextual adaptability. We then trace them through thirteen design principles to implementation choices. The core of the system is a simple while-loop that calls the model, runs tools, and repeats. Most of the code, however, lives in the systems around this loop: a permission system with seven modes and an ML-based classifier, a five-layer compaction pipeline for context management, four extensibility mechanisms (MCP, plugins, skills, and hooks), a subagent delegation and orchestration mechanism, and append-oriented session storage. Comparisons with OpenClaw and Hermes Agent show that the same design questions produce different answers across three deployment contexts. Claude Code emphasizes per-action safety, OpenClaw emphasizes perimeter-level access, and Hermes renders per-action approvals across many surfaces. At the runtime layer, Claude Code uses a single CLI loop, OpenClaw embeds the runtime within a gateway control plane, and Hermes uses one process whose role is set by its entry point. At the context and extension layer, Claude Code extends the context window, OpenClaw registers gateway-wide capabilities, and Hermes provides pluggable memory and model backends. We finally identify six open design directions for future agent systems, grounded in recent empirical, architectural, and policy literature.

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

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, \textsc{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.

Multi-agent systems composed of large generative models are rapidly moving from laboratory prototypes to real-world deployments, where they jointly plan, negotiate, and allocate shared resources to solve complex tasks. While such systems promise unprecedented scalability and autonomy, their collective interaction also gives rise to failure modes that cannot be reduced to individual agents. Understanding these emergent risks is therefore critical. Here, we present a pioneer study of such emergent multi-agent risk in workflows that involve competition over shared resources (e.g., computing resources or market share), sequential handoff collaboration (where downstream agents see only predecessor outputs), collective decision aggregation, and others. Across these settings, we observe that such group behaviors arise frequently across repeated trials and a wide range of interaction conditions, rather than as rare or pathological cases. In particular, phenomena such as collusion-like coordination and conformity emerge with non-trivial frequency under realistic resource constraints, communication protocols, and role assignments, mirroring well-known pathologies in human societies despite no explicit instruction. Moreover, these risks cannot be prevented by existing agent-level safeguards alone. These findings expose the dark side of intelligent multi-agent systems: a social intelligence risk where agent collectives, despite no instruction to do so, spontaneously reproduce familiar failure patterns from human societies.