SkillSpector
NVIDIA's scanner that checks an agent skill or MCP server for prompt injection, data exfiltration and other risks before you install it.
Install
-
CLI
uv tool install git+https://github.com/NVIDIA/skillspector.gitPin to a reviewed commit by appending @<sha>.
-
Scan a skill
skillspector scan ./my-skill/ --no-llm -
Claude Code skill
Copy skills/skill-inspector from the repo into ~/.claude/skills/ for a scan plus a read-the-source review.
The tool this catalog uses to vet everything in it. It runs 71 checks across 17 categories — prompt injection, exfiltration, privilege escalation, MCP tool poisoning and more — and scores the result from 0 to 100.
How to read a scan
Treat the score as a reason to look, not a verdict. It flags patterns: a skill that
legitimately runs Python or calls an API can score 100. The bundled skill-inspector skill
pairs the scan with a read of the source behind every serious finding, and ends in
APPROVE, CAUTION or REJECT.
Good companions
- gitleaks for secrets committed in a skill’s repo.
- osv-scanner for known vulnerabilities in bundled dependencies.
- OpenSSF Scorecard for how well the source repository itself is maintained.
- Publisher
- NVIDIA
- Works with
- Any agent, Claude Code, Codex, Gemini
- Categories
- Security & trust, Agent building
- License
- Apache-2.0
- Cost
- Free