Fadé Adeniyi
Friday 18 September 2026
NVIDIA checked over 42,000 real agent skills. 1 in 20 showed signs of being malicious. Here's their free scanner, and exactly how to point it at your own skills folder.
A Claude skill isn't just a harmless text file. It runs with whatever access you've given it, which can include your environment variables, your API keys and your local client folders. The dangerous instructions can be hidden in text you never see when you open it.
NVIDIA studied 42,447 real agent skills pulled from major marketplaces. 26.1% had at least one security vulnerability. 5.2% showed patterns strongly suggestive of malicious intent. This is for anyone who's installed more than one Claude Code skill without checking what it does first.
NVIDIA built a free scanner for exactly this.
1. Get SkillSpector. It's free and open source, on GitHub under NVIDIA's own account: github.com/NVIDIA/SkillSpector.
2. Clone it and install: git clone https://github.com/NVIDIA/SkillSpector.git && cd SkillSpector && pip install -e .
3. Point it at a single skill: skillspector scan ~/.claude/skills/some-skill. Or a whole folder at once: skillspector scan ~/.claude/skills/.
4. Read the risk score. It's out of 100, with a plain answer on whether it's safe to install.
5. Optionally, wire it in as an MCP so Claude scans anything new automatically, before it ever runs.
It checks 64 real patterns across 16 categories, including credential exfiltration, hidden remote-script execution and cron jobs that keep running after you've closed the terminal.
Because this isn't a random tool making the claim. The 42,447-skill study is NVIDIA's own published research, not a rumour. NVIDIA uses the scanner on itself, and nothing gets published to their own skills catalogue until it passes this exact scan. And it's genuinely free: open source under an Apache 2.0 licence, with no cost to run it on your own folder.
A skill looking clean when you open it means nothing. Malicious instructions can hide in invisible text, comments or encoded scripts, precisely so a manual read-through won't catch them. That's the whole reason an automated scanner exists.
A clean scan today also doesn't mean a skill stays clean if it auto-updates later. Re-scan periodically.
Clean on the page isn't clean underneath.
A scan is a risk score, not a guarantee. Nothing here proves a skill is safe, only that it didn't match the 64 known patterns. The figures are the real ones and worth quoting exactly: 5.2% malicious, 26.1% vulnerable, out of 42,447 real skills.
So point SkillSpector at your existing skills folder before you install a single new one this week. One in twenty skills people install is showing signs of malicious intent, and you don't have to guess which ones.
github.com/NVIDIA/SkillSpector, Apache 2.0 licence, real published research from NVIDIA.
The Audacity to Build is my community for people doing exactly that. It is opening soon. Join the waitlist and you are first in when it does.
If this was useful, follow me. I share what I am actually building, the tools that survive my own testing, and the ones I quietly cancel. New plays every week.
Pop in your name and email and the rest of the guide is yours, along with the occasional new guide when I publish one. No spam, ever.