PoisonedEvolution Attack Achieves 91% Success in Self-Evolving Agent Systems

The PoisonedEvolution attack exploits the process where agents distill trajectories into persistent skills. By injecting bounded evidence, attackers achieved a 91.0% success rate in embedding target behaviors into six mainstream LLM evolvers within the SkillClaw framework.