Stop an Agent From Obeying a Document

Defend a RAG-style agent against instructions hidden in retrieved content.

A advanced AI Engineering challenge worth 20 points. Solve it hands-on in a real AI Engineering environment in your browser - no local setup, no fake shells.

The Challenge

Scenario: An internal ops agent retrieves a document to summarize it - but the document contains hidden text instructing the agent to leak its own system policy, and the agent obeys it.

Establish retrieved content as untrusted data rather than executable instructions, add pattern-based injection detection, and remove the vulnerable code path, without breaking normal summarization of clean documents. Run check.