← Videos
Can Docs Trick AI - Injection

AI from the Ground Up8 Nov 2026 · 4:48

Can Docs Trick AI - Injection

A document contains the answer you need—and an instruction telling the assistant to ignore its task. Which part should the assistant follow?

From the episode description

A document contains the answer you need—and an instruction telling the assistant to ignore its task. Which part should the assistant follow? We use a harmless fictional example to explain indirect prompt injection: a workshop note includes a sentence telling the assistant to say “BANANA.” The question is not whether that word is dangerous. It is whether source content can redirect an authorized task. In this episode: • How instructions can arrive inside documents or tool results. • Why useful evidence and untrusted commands can share the same passage. • How answer quality and tool permissions require separate checks. • Why read access should not silently become permission to send or edit. • How to compare an original note with an injected version. Workshop Notes is fictional, and the expected outcome is a test specification, not proof that a particular model resists the attempt. No live service is attacked, and no control is presented as a perfect defense. Keep two questions in view: what does the source support, and who authorized the action? That distinction helps you inspect a system without mistaking a plausible response for an enforced permission boundary. ResearchNode: AI from the ground up • Episode 09 of 10. Previous: MCP Explained: How AI Connects to Your Knowledge Next: How Do You Know Your AI Assistant Works? Presented by Michelle, an AI-generated presenter.

Descriptions and chapter markers are supplied by the channel. They are not a transcript or an independent check of every claim in the video.