CLAIM-2026-10-13-04

The paper reports that guard models miss many deceptive requests and that Self-Reminder reduces but does not eliminate deceptive responses.

Why it matters

These are the two defences most teams already run, so a practitioner cannot treat either as settling the problem.