One Lesson, Three Ways

Why retrieved text cannot authorize an AI agent

Data is not permission

An agent may read a document to answer a question. That document can contain hostile instructions. The agent’s allowed actions must come from trusted system rules and explicit tool permissions, not from the document itself.

1. Identify the source of the instruction.
2. Check whether that source is authorized to request the action.
3. Limit tool permissions to the task.
4. Escalate consequential actions for human approval.

This is a written explanation. An instructor-recorded lesson can be added when approved footage is available.