SkillSpector

Installing an agent skill is handing it the keys. A skill is Markdown plus scripts that run with your shell privileges, and most teams install them without a second look. SkillSpector, from NVIDIA, scans a skill before it is installed and checks it against 71 vulnerability patterns across 17 categories, from prompt injection and data exfiltration to dependency CVEs. It returns a risk score and a verdict, and it can run as an MCP server so agents scan before installing. In the research behind it, 26.1 percent of 31,132 skills analysed had vulnerabilities. Scan first, install second.

What you get

  • A 0-100 risk score with severity labels and clear recommendations, reported in terminal, JSON, Markdown, or SARIF output
  • Live CVE lookups against OSV.dev for dependency checks, with an automatic offline fallback
  • A Docker image if you want to run it without installing Python, and baseline suppression so re-scans surface only new findings
  • Install with uv tool install --python 3.13 git+https://github.com/NVIDIA/skillspector.git, Python 3.12 or 3.13 required

A skill runs with your shell privileges. Treat installing one like installing software from a stranger: scan it first, and only then decide whether it earns the keys.

Repo: https://github.com/NVIDIA/SkillSpector

Quick poll

What would you use SkillSpector for first?

  • Scanning skills before I install them
  • Gating skill installs in CI
  • Running it as an MCP server for my agents
  • Watching it for now

Vote in the comments.

If you found SkillSpector interesting, I write about tools like this every week in The AI-Native Engineering Leader, what actually works for engineering leaders building with AI agents, and what to skip.