Compositional Safety Failures in Harness Evolution: Identification and Runtime Monitoring
Self-evolving agent harnesses continually update persistent components such as memory, prompts, skills, and tools. We call this process harness evolution. However, such evolution could introduce unexpected safety risks. Existing work studies harness misevolution and validates candidate harnesses or attributed individua...
Zhi-Xiang Zhang, Ze-Sen Liu, Wai-Ip Lai et al.
· 0 citations