SkillCloak lets malicious AI coding skills dodge every static scanner
Researchers at HKUST show that the SkillCloak framework repackages malicious LLM agent skills with self‑extracting packing, achieving >90% bypass rates across eight scanners, exposing the weakness of appearance‑based defenses. Their follow‑up system, SkillDetonate, restores detection by sandboxing and data‑flow analysis, catching 97% of attacks.