Who it is forA developer about to install a community skill into Claude Code or Codex.
The problemSkills run with implicit trust; the README cites research where 26.1% of analyzed skills had vulnerabilities and 5.2% showed likely malicious intent.
The outcomeA 0 to 100 risk score with severity labels and findings, so you decide whether to install.
What you need first
- Python with uv or pip, or Docker (no Python required)
- Optional: an LLM API key in .env for semantic analysis (default config targets DeepSeek)
- Optional MCP extra for skillspector mcp
Steps
- Install with
uv tool install git+https://github.com/NVIDIA/skillspector.git - Download or clone the skill you want to check
- Scan it without LLM costs using Docker:
docker run --rm -v "$PWD:/scan" skillspector scan ./my-skill/ --no-llm - Read the risk score, severity labels and recommendations
- Install the skill only if the findings are acceptable
Prompt for your agent
Use NVIDIA/SkillSpector to scan the skill at <path or repo> before I install it. Run static analysis only, then summarize the risk score and each finding in plain words, and recommend install, review or reject.
Caveat. A clean scan lowers risk but is not a guarantee; static analysis cannot prove a skill is safe.
Use this in your AI agent
Ask your agent
Using the Deduped MCP, get the setup steps for NVIDIA/SkillSpector and "Scan an agent skill for prompt injection and exfiltration before installing"
Add Deduped to Claude Code once
claude mcp add --transport http deduped https://dedupedai.com/mcp
Every Monday: the 5 hottest repos and exactly how to use them. Free, one email a week, unsubscribe anytime.
Get them freeShare: X LinkedIn Hacker News Email Copy link
Third-party repos we don't own. AI-written use cases can be wrong. Disclaimer