Policy leak finderAI agents work under your permission policy · an AI reviewer flags what goes wrong · code confirms it

…

The setup

How violations are flagged
  1. After every conversation, an AI reviewer reads every question, answer and message, and reports anything an agent learned or changed that it must never.
  2. Code confirms each finding against the real data: is it true, and can it really be worked out from answers that agent received? Unconfirmed claims are rejected.
  3. Simple code checks run as a backstop, and show anything the AI reviewer missed.

Permission policy click any cell for details

allowednot allowedpolicy says nothingMUST NEVER… protected by the goalviolated anyway

Agent activityeach square is one question an agent asked the system · click one