A retrieved page instructs the assistant to ignore its rules and disclose private information. Which approach best fits?

Published
  1. Use live retrieval or tools for current inventory rather than assuming tuning updates facts continuously.
  2. Require an evidence threshold and an explicit abstention or escalation path.
  3. Create representative evaluation cases including edge cases and failure scenarios.
  4. Treat retrieved instructions as untrusted content and enforce application-level access controls.

Comments — 0