AgentHarm turns agent misuse into a concrete safety benchmark
AgentHarm measures whether LLM agents refuse malicious multi-step tool-use requests and whether jailbreaks preserve enough capability to complete harmful tasks.
原始内容为英文；当前页面提供中文导航与来源说明，具体事实请以原文为准。
为什么值得关注
AgentHarm reframes safety testing around agents that use external tools and execute multi-stage tasks, rather than simple chatbot refusals. The benchmark includes 110 explicitly malicious agent tasks, expanded to 440 with augmentations, across 11 harm categories including fraud, cybercrime, and harassment. For agent builders, the signal is practical: safety evals need to test refusal, jailbreak resistance, and whether compromised agents still retain the planning and tool-use ability needed to complete harmful workflows.
可执行摘要
Use AgentHarm-style evals to test malicious task refusal, jailbreak robustness, multi-step tool-use capability retention, and harm-category coverage.
- Agent 实用度
- 93/100
- 可信度
- 88%
- 机器格式
- JSON + Markdown