Scan a Claude skill before you install it: NVIDIA's free SkillSpector

1 in 4 Claude skills in one big study had a security hole. SkillSpector is NVIDIA's free, open-source scanner that reads an agent skill before you install it and tells you SAFE, CAUTION or DO NOT INSTALL, with the file and line of each problem. This is the full version of the SCAN guide from my Reel. Facts checked on 20 Oct 2026 against the GitHub README and NVIDIA's docs.

01What it is

SkillSpector (github.com/NVIDIA/SkillSpector, Apache-2.0) is a security scanner for agent skills used by Claude Code, Codex, Gemini CLI and MCP servers. It scans a skill folder, a single SKILL.md, a zip file, a URL or a Git repository. It is also the scanning step inside NVIDIA's Verified Skills pipeline.

02Why it matters

Skills run with a lot of implicit trust: they can include scripts, instructions and dependencies your agent will follow. In the 31,132-skill research subset the README cites, 26.1% of skills contained vulnerabilities and 5.2% showed likely malicious intent. GitHub stars tell you a skill is popular, not that someone read its code.

03What it checks

Attack patterns across 17 categories, including prompt injection, data exfiltration, privilege escalation, supply-chain risk, excessive agency, memory poisoning, tool misuse, dangerous code and MCP tool poisoning. (The README lists 71 patterns; NVIDIA's docs page currently says 68.) It runs fast static analysis first and an optional LLM review on top.

04Step by step

Install it (needs Python 3.12+ and uv): uv tool install git+https://github.com/NVIDIA/skillspector.git
Scan the skill before you install it. Folder: skillspector scan ./my-skill/ . Zip: skillspector scan ./my-skill.zip . GitHub: skillspector scan https://github.com/user/my-skill
Fast static-only scan, no API key: add --no-llm.
For the LLM stage, set a provider, for example SKILLSPECTOR_PROVIDER=anthropic and ANTHROPIC_API_KEY.
Save a report: --format json --output report.json (also markdown and SARIF).
No Python? Build the Docker image (docker build -t skillspector .) and run: docker run --rm -v "$PWD:/scan" skillspector scan ./my-skill/ --no-llm

Install and run a fast scan

uv tool install git+https://github.com/NVIDIA/skillspector.git
skillspector scan ./my-skill/ --no-llm

05Reading the result

You get a risk score from 0 to 100: 0-20 SAFE, 21-50 CAUTION, 51-100 DO NOT INSTALL. Each issue shows its severity, the file and line (for example scripts/sync.py:23), a confidence and an explanation. The CLI exit code is 0 for SAFE or CAUTION and 1 above 50, so you can use it as an install gate in CI.

06Good to know

A low score from a static-only (--no-llm) scan is not the same as a full review; the JSON output says which mode ran.
CAUTION exits with 0 by default; add --fail-on-findings if you want any finding to block.
Known false positives can be accepted with a baseline file so re-scans only show new issues.