SkillHarm: Lifecycle-Aware Skill-Based Attacks on AI Agents via Automated Construction

arXiv·high signalXBlueskyLinkedInCopy link

Introduces SkillHarm, a framework that automatically constructs adversarial agent skills targeting every phase of the skill lifecycle — creation, distribution, deployment, and execution. The paper demonstrates that third-party skills represent a privileged and under-secured attack surface, as agents implicitly trust and execute skill instructions. Builders deploying agent skill registries need to treat skills as untrusted code with sandbox isolation and integrity verification.