New SkillCloak Technique Enables Malicious AI Skills to Evade Detection

A new technique called SkillCloak allows malicious AI agent skills to bypass existing static skill scanners over 90% of the time. These skills can steal credentials, exfiltrate source code, and install backdoors, posing a significant risk to AI coding assistants like Claude Code and OpenAI Codex. Current defenses, primarily relying on static pattern matching, are ineffective against adaptive evasions. Researchers emphasize that runtime behavior auditing offers better resilience against these threats. The scope of impact includes a wide range of AI tools that utilize third-party agent skills, creating a substantial software supply-chain attack surface. As of now, the threat remains active, with no confirmed mitigation strategies reported.

Key Points: • SkillCloak technique bypasses static scanners over 90% of the time. • Malicious agent skills can steal credentials and install backdoors. • Runtime behavior auditing is more effective than static scanning.