A tool response can mix useful facts with hostile instructions. It might return a real ticket summary followed by a request to upload credentials for verification. Preserve the useful data while rejecting that attempted change of authority. In this lesson you mark tool output as data, block credential transmission and record the source of the injected instruction. The exercise includes a response that claims administrative status. That claim is part of the untrusted content. Review any resulting tool call against the original task and explicit grants. An injection warning in the final summary does not repair a credential already sent in an earlier argument.
A worked artifact for a separate fictional task. Compare its actions with the contract.