New 'SkillCloak' Technique Lets Malicious AI Agent Skills Evade Scanners

Researchers at HKUST developed SkillCloak, a method that packs malicious AI agent skills into self-extracting archives, evading static scanners over 90% of the time. The technique exploits how scanners analyze skill code, leaving malware functional. The team also built a runtime checker that catches most attacks, highlighting a growing security gap in AI agent ecosystems.