RedEvoAgent: Automatic Red-Teaming Agent with Experience-Driven Skill Evolution
RedEvoAgent turns prior attack runs into a shorter evolving skill instead of replaying full trajectories.
The paper frames the risk around LLM agents that can use tools and change persistent state, not just produce unsafe text. Its proposed black-box red-teaming agent profiles tool effectiveness, attributes which tool helped an attack, and keeps only skill updates that improve validation performance. The authors report gains over fixed and agentic baselines across benchmarks, target models, and execution harnesses. ArXiv · AI/CL/LG's note
The paper frames the risk around LLM agents that can use tools and change persistent state, not just produce unsafe text. Its proposed black-box red-teaming agent profiles tool effectiveness, attributes which tool helped an attack, and keeps only skill updates that improve validation performance. The authors report gains over fixed and agentic baselines across benchmarks, target models, and execution harnesses. ArXiv · AI/CL/LG's note
score 4