SkillCloak Lets Malicious AI Agent Skills Evade Static Scanners with Self-Extracting Packing

Researchers show that malicious "skills" for AI coding agents can be cloaked to bypass static scanners (via look-alike characters, split commands, and packing payloads into ignored directories like .git/), allowing credential theft, source-code exfiltration, or backdoors; a runtime sandbox detector (SKILLDETONATE) catches most attacks but is slower. Real-world scans and reports (Bitdefender, Koi Security, Unit 42) already find many malicious skills on marketplaces, so defenders should treat a "passed scan" as a starting point: monitor runtime behavior, flag large/opaque files in ignored folders, re-check hashes before execution, and apply least-privilege to agents and vetted sources.