NVIDIA/SkillSpector
SkillSpector scans agent skills for security risks before installation, with static checks, optional LLM analysis, and reports that can feed CI. Its cited research subset found vulnerabilities in 26.1% of skills.
SkillSpector checks skills from repositories, URLs, archives, directories, or files using **71 vulnerability patterns across 17 categories**. It combines static analysis with optional LLM review and can emit SARIF, JSON, Markdown, or terminal reports. The cited dataset subset contained 31,132 skills; **26.1% had vulnerabilities** and 5.2% showed likely malicious intent. Add a scan before installing third-party skills, or use SARIF and `--fail-on-incomplete` to support CI gates. Baselines can suppress findings already reviewed so later scans surface new ones. A scanner report is a review aid, not proof a skill is safe. Findings can be suppressed, and caller exclusions make coverage partial and prevent a SAFE recommendation. LLM analysis is optional and requires provider configuration.
No separate implication has been recorded yet.
SkillSpector checks skills from repositories, URLs, archives, directories, or files using **71 vulnerability patterns across 17 categories**. It combines static analysis with optional LLM review and can emit SARIF, JSON, Markdown, or terminal reports. The cited dataset subset contained 31,132 skills; **26.1% had vulnerabilities** and 5.2% showed likely malicious intent. Add a scan before installing third-party skills, or use SARIF and `--fail-on-incomplete` to support CI gates. Baselines can suppress findings already reviewed so later scans surface new ones. A scanner report is a review aid, not proof a skill is safe. Findings can be suppressed, and caller exclusions make coverage partial and prevent a SAFE recommendation. LLM analysis is optional and requires provider configuration.
This makes third-party skill review a repeatable scanning step, with CI output and baselines to track newly surfaced findings. Its dataset figures give scale to the exposure concern, while the caveats matter operationally: a clean report is not a safety guarantee, and exclusions can leave coverage incomplete.