NVIDIA SkillSpector Scans AI Agent Skills for Hidden Threats

Installing a random Claude Code skill from GitHub is a quick way to make your agent more useful. It is also a quick way to hand over your API keys. NVIDIA's SkillSpector is a free, open-source scanner built to catch that before it happens.

It reads a skill and flags the parts that could bite you, so you know what you are letting into your coding setup.

What Happened

SkillSpector is an open-source security scanner for AI agent skills. Instead of trusting a skill's description, you run it through the scanner first and look at what it finds. It checks for four kinds of trouble:

• Prompt injection attacks
• Credential theft and data exfiltration risks
• Dangerous code and privilege escalation
• Malicious dependencies and supply-chain threats

It supports skills used with Claude Code, Codex, and other AI agent workflows. The point is to make it a security checkpoint, the thing you pass a third-party skill through before it reaches your machine.

Why It Matters

A skill is instructions plus code that your agent agrees to follow. Some of them carry hidden instructions or behavior that goes looking for sensitive data and tries to send it somewhere else. That means your credentials, and anything else sitting in your projects, can leak from a skill you installed once and forgot about.

The scanner does not decide for you. It shows you what is inside, and the habit of checking becomes the real protection. If you are piling on skills to make your agent more powerful, it is worth looking at what each one actually does before you install it. The source says the GitHub link is in its comments, so check there for the project itself.