31,132 AI Skills Scanned: a Quarter Have Holes, 5% Don't Look Benign
NVIDIA's open-source SkillSpector lays bare the security problems with AI agent skills: of 31,132 skills, 26.1% contain vulnerabilities and 5.2% show clear malicious intent. Scan before you install — but don't treat the scanner as insurance.
Installing an AI agent skill is a lot like downloading a plugin from a forum. The difference is that it often runs directly inside Claude Code, Codex CLI, or Gemini CLI, holding your permissions, able to read project files and touch environment variables. Many people hit install without ever opening SKILL.md.
NVIDIA's open-source SkillSpector puts a set of numbers right on its homepage: of 31,132 skills, 26.1% have at least one vulnerability, and 5.2% exhibit clear malicious intent. The data comes from the paper Agent Skills in the Wild, by Yi Liu and others. Someone in the X comments put it bluntly: 5% malicious is no longer within the margin of error — skills look like they're walking the same road npm did.
But that 26.1% figure needs to be broken down. As one commenter noted, 12.8% of the skills in the paper have only low-severity issues, such as unpinned dependency versions or overly broad permission declarations. The real trouble is skills that ship executable scripts: they are 2.12 times more likely to have vulnerabilities than instruction-only skills. A skill that only writes prompts carries less risk; once it includes scripts/, it can do things on your machine.
SkillSpector can scan local directories, a single SKILL.md, Git repositories, and zip packages. It ships with 71 vulnerability patterns, covering categories such as prompt injection, data exfiltration, privilege escalation, supply chain, and overly broad MCP permissions. By default it runs static analysis only, with LLM semantic analysis optional; output supports terminal, JSON, Markdown, and SARIF. It can also be invoked by Claude Code as an MCP server, and its exit codes can hook into CI to gate skill installs in the pipeline.
But don't treat the scanner as insurance. It doesn't execute the skill, and it doesn't isolate the host. Static analysis can catch obvious curl | bash, environment-variable harvesting, and exec(), but it can't catch commands assembled at runtime. One commenter's approach is more practical: read it before installing, pin to a specific commit, run the agent as a low-privilege user, and don't leave ~/.ssh and .env exposed. A scanner lowers your odds of stepping on a landmine, but permission management is the real baseline.
发布时间: 2026-11-19 00:28