Prompt injection

Untrusted content attempts to redirect a model away from the application's intended instructions.

Example

A retrieved page tells the assistant to ignore its instructions and leak data.

Treat retrieved text as data, not authority to change permissions.

Reference definition