Prompt injection.

Prompt injection is when malicious instructions get smuggled into your AI agent through something you installed: a Skills, a tool, a Plugin. The agent reads those hidden instructions as if they came from you and acts on them. A tool that looks helpful could quietly tell your agent to leak data or do something you never asked for.

It's the number one critical vulnerability in AI agent systems, top of OWASP's 2025 list for LLM apps. It gets more dangerous the more tools you hand your agent, because every new skill or plugin is another door someone could have tampered with.

So treat anything you install like code you're running on your own machine, because that's what it is. Check where it came from before you give your agent access.