Prompt injection
Definition: Prompt injection is an attack where malicious content (in a page or document) hijacks a model's instructions to make it take unintended actions.
It's a major risk for tool-connected AI agents. The defense: treat external content as data, never as commands.
Frequently asked questions
What is a prompt injection attack?
Booby-trapped content (in a web page, email or document) that hijacks a model's instructions to make it take unintended actions: exfiltrate data, call a tool, ignore its rules.
How do you defend against it?
Treat all external content as untrusted data, never as commands; scope tool permissions tightly; require human approval for sensitive actions; filter both inputs and outputs.