Before You Install That AI Skill, Scan It

NVIDIA built a free scanner that checks AI agent skills for hidden dangers before you run them — because one in four have real vulnerabilities.

Something worth paying attention to

There's a quiet problem growing in the world of AI agents. People are building small add-ons — called skills — that give AI assistants new abilities: booking meetings, reading files, sending emails, talking to other software. These skills get shared publicly, and anyone can install them.

The trouble? A study cited by NVIDIA found that more than a quarter of these publicly available skills contain at least one security flaw. And about 5% appear to be genuinely malicious — designed to quietly steal things like passwords or account credentials while pretending to do something harmless.

NVIDIA's response is a free tool called SkillSpector. You point it at any skill — whether it's a link, a file, or a folder — and it returns a score from 0 to 100. Low score, you're probably fine. High score, walk away.

It checks for things like: does this skill claim to sync your calendar but actually try to read your saved passwords? Does it ask for more access than it should? It looks across 64 known risk patterns — and does the basic check instantly, with no account required.

This matters even if you're not technical. If you're using an AI assistant and your team is adding skills or plugins to it, someone is choosing what to install. This tool is how you check before you trust.

Words worth knowing

AI skill / plugin — A small add-on that gives your AI assistant a new ability, like reading your inbox or checking inventory.

Prompt injection — A trick where a skill secretly rewrites the AI's instructions to behave differently than you expect.

Credential exfiltration — A fancy way of saying: stealing your passwords or account keys without you noticing.

MCP — A standard way for AI agents to connect to external tools and services. Think of it as a universal plug socket for AI.

If your team is experimenting with AI agents and adding community-built skills to them, ask whoever is doing the installing: are we checking these before we run them? That question alone is worth asking at your next meeting.