Researchers from Hong Kong University demonstrated that AI skill scanners can be bypassed by malicious agents using a technique called SKILLCLOAK. This method rewrites skills to evade detection, highlighting significant security risks for AI coding agents.
The SKILLCLOAK tool rewrites malicious skills to appear clean while maintaining their functionality. It incorporates two primary methods: character swapping to confuse scanners and self-extracting packing that conceals payloads within directories scanners ignore.
The researchers tested SKILLCLOAK against eight scanners with 1,613 real malicious skills, achieving evasion rates exceeding 90%. The success of this tool demonstrates the inadequacy of current skill scanners in detecting sophisticated threats.
The findings raise critical concerns regarding the use of third-party skills by AI coding agents, which often operate with extensive access to user files. Malicious skills could potentially execute harmful actions such as credential theft or backdoor installation.
Existing defenses rely heavily on skill scanners that failed to adequately protect against the techniques demonstrated by SKILLCLOAK. The study suggests that a new runtime checker, developed by the same research team, could detect malicious skills that evade traditional scanners.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
Researchers from Hong Kong University demonstrated that AI skill scanners can be bypassed by malicious agents using a technique called SKILLCLOAK. This method rewrites skills to evade detection, highlighting significant security risks for AI coding agents.