SkillSpector: scan Claude Code skills before you install them

Why it matters

• Researchers found that 26.1% of 31,132 marketplace skills had at least one vulnerability and 5.2% showed high-severity patterns strongly suggesting malicious intent (Liu et al., arXiv 2601.10338, Jan 2026).
• The follow-up study confirmed 157 malicious skills out of 98,380 (about 0.16%), all removed (arXiv 2602.06547).

Install (tested 24 Oct 2026, v2.12.0)

1. 01Install uv, then SkillSpector

   terminal

   COPY

   uv tool install git+https://github.com/NVIDIA/skillspector.git

2. 02Only Python 3.14 on your machine?

   Use:

   terminal

   COPY

   uv tool install --python 3.12 git+https://github.com/NVIDIA/skillspector.git

3. 03It is not on PyPI

   Treat any pip install skillspector package as a possible typosquat.

Scan (offline, nothing leaves your machine)

1. 01One skill

   terminal

   COPY

   skillspector scan ./my-skill --no-llm

2. 02Everything you already have

   One level deep: each subfolder with a SKILL.md.

   terminal

   COPY

   skillspector scan ~/.claude/skills --recursive --no-llm

3. 03Before installing

   Point it at the GitHub link or the zip.

Read the score

Our own offline test on the tool's fixtures: safe_skill 0/100 (no findings); malicious_skill 100/100 with 7 findings (harmful instruction, environment-variable harvesting, external transmission).

A low score means no known red flags were found, not that a skill is safe.

What it cannot see

• --no-llm is static only: it caught every code-level attack in our test but missed natural-language attacks (the LLM review stage exists for those, and it sends file contents to the provider you choose).
• It cannot read text in images, binaries or runtime behaviour.

The safe rule

Scan every skill before you install it, keep the scan offline with --no-llm, and treat a low score as "no red flags found", not "safe".

Official links

• NVIDIA SkillSpector (Apache-2.0)github.com/NVIDIA/SkillSpector ↗
• NVIDIA docs: Scanning agent skillsdocs.nvidia.com/skills/scanning-agent-skills ↗
• Liu et al., agent skills security studyarxiv.org/abs/2601.10338 ↗

Sources checked

The facts on this page were checked against the sources below, and the scan results come from our own offline test of v2.12.0 on 24 October 2026. SkillSpector changes fast, so treat its README as the final word.

• SkillSpector: NVIDIA, Apache-2.0, 17 categories, 0-100 risk score, scans folders, GitHub links, URLs and zips; not on PyPI. Source ↗
• What SkillSpector checks: prompt injection, data exfiltration, tool misuse among 17 categories. Source ↗
• 26.1% of 31,132 marketplace skills had at least one vulnerability; 5.2% high-severity patterns suggesting malicious intent. Source ↗
• Follow-up: 157 confirmed malicious skills out of 98,380, all removed. Source ↗
• Our offline test (v2.12.0, 24 Oct 2026): safe_skill 0/100 SAFE, malicious_skill 100/100 DO NOT INSTALL. Source ↗