// Skill Misevolution in Self-Improving LLM Agents //

Self-improving agents write their successes down as reusable skills. An unsafe success becomes reusable policy long after the input that triggered it is gone.

SkillMisevo-Gym versions skill state across agent frameworks so risk can be attributed separately to authoring, retrieval, and later execution.

Across 25 agent-method configurations covering 525 tasks each, all 21 evolved configurations author unsafe artifacts. Only 15 produce harm in a fresh session, which means authoring risk and execution risk are separate problems.

Three malicious tasks raise carryover attack success from 16.0% to 35.3%.

Their SafeEvolve wrapper cuts unsafe retrieval by 26.7 points and fresh-session harm by 17.3 points while benign utility moves 0.4 points.

Paper: https://arxiv.org/abs/2608.12851

Track more trending AI papers in our academy: https://academy.dair.ai/