Researchers went through 42,447 Claude skills from two public marketplaces, and about 1 in 20 showed signs of being malicious. A skill can use everything you can: your files, your work folders and any keys you’ve got saved.
This guide sets up SkillSpector, NVIDIA’s free scanner, and then tells Claude to scan every skill before it installs one, so you’re not relying on remembering.
Before you start
What you need: Claude Code, and a Mac or Linux terminal. Windows works through WSL.
How long: about 10 minutes.
Tech skills: copy and paste into a terminal. That’s it.
Free or paid: SkillSpector is free and open source (Apache 2.0). Claude Code needs a paid Claude plan.
What the score means
SkillSpector checks a skill against 71 warning signs and gives it a risk score out of 100.
| Score | Rating | What to do |
|---|---|---|
| 0 to 20 | Low | Safe to install |
| 21 to 50 | Medium | Read the findings first |
| 51 to 80 | High | Do not install |
| 81 to 100 | Critical | Do not install |
The rule: anything over 50, don’t install it.
Step 1: Install uv
Time: 1 minute.
uv is a small tool that installs Python apps cleanly. If you already have it, skip this step.
curl -LsSf https://astral.sh/uv/install.sh | shClose the terminal and open a new one so it picks uv up.
Step 2: Install SkillSpector
Time: 2 minutes.
This installs the scanner with the part that lets Claude use it.
uv tool install 'skillspector[mcp] @ git+https://github.com/NVIDIA/skillspector.git'You’ll know it worked when
skillspector --helpprints a list of commands.
Step 3: Scan one skill by hand
Time: 2 minutes.
Try it on a skill you already have, or on any skill’s GitHub link.
skillspector scan ~/.claude/skills/[skill-folder-name] --no-llmskillspector scan https://github.com/[owner]/[repo] --no-llm--no-llm runs the full rule check with no API key needed. SkillSpector can also run a deeper AI review on top, which needs an NVIDIA key. The rule check is plenty to start with.
You’ll know it worked when you get a score out of 100 and a list of anything it flagged.
Step 4: Connect it to Claude Code
Time: 1 minute.
claude mcp add skillspector -- skillspector mcpThis gives Claude a scan tool it can call itself.
You’ll know it worked when you open Claude Code, type
/mcp, and skillspector is in the list.
Step 5: Make Claude scan every new skill
Time: 2 minutes.
On its own, Claude only scans when you ask. This step makes it the default. Open your global instructions file at ~/.claude/CLAUDE.md (create it if it isn’t there) and paste this at the bottom:
Before installing any skill, plugin or MCP server, scan it first with the skillspector scan_skill tool. Use use_llm=false unless I've set up an NVIDIA key. Show me the risk score and anything it flagged.
- 0 to 20: go ahead and install.
- 21 to 50: show me the findings and ask me before installing.
- Over 50: do not install it. Tell me why.
This applies even if I ask you to install something quickly.This is an instruction, not a lock. Claude follows it, but it isn’t a hard block. If you ever install a skill by hand outside Claude, scan it yourself with step 3.
You’ll know it worked when you ask Claude Code to install a skill and it scans it and shows you the score before doing anything.
The final check
-
skillspector --helpworks - I’ve scanned one skill by hand and got a score
- skillspector shows up in
/mcpinside Claude Code - The scan rule is in
~/.claude/CLAUDE.md - Claude scanned a skill before installing it
If something looks different
“command not found: skillspector”. Open a new terminal window. If it still fails, run uv tool update-shell, then open a new one again.
It asks for an API key. That’s the optional AI review. Add --no-llm when you scan by hand, and keep use_llm=false in the CLAUDE.md line.
Claude installed a skill without scanning it. Check the line is in ~/.claude/CLAUDE.md (the global one, not a project file), then start a new Claude Code session so it reloads.
A skill I trust scored over 50. Read what it flagged. Some flags have innocent reasons, but don’t install it until you understand every one.
Where this comes from
- SkillSpector on GitHub: github.com/NVIDIA/SkillSpector
- The study of 42,447 skills: Agent Skills in the Wild (Liu et al., 2026)
- NVIDIA runs the same scan on its own skills before publishing them: NVIDIA developer blog