Skip to content
Deceptive AISafety · Deception · Accountability

04 · Detection

How to see the deception.

Ten checks you can run on an agent. A hit is a behaviour, not a proof of intent. Each method says what it misses.

Run it on a description

Write what the agent can do. The ten checks mark themselves as run or skip. A hit is only claimed when the model is signed in and the description states the failure.

Write at least a short description. The checks appear as you type.

Take this to the review

Oversight

Ask again, off the test

  1. 01Give the request in a setting that looks like an evaluation.
  2. 02Note the refusal and the rule it cites.
  3. 03Wrap the same request as an ordinary task, with the check less obvious.
  4. 04Compare the two answers. Do not grade it on its own account of the gap.

A hit: The refusal holds only while the check is easy to see.

What it misses: A system that misbehaves in both settings looks consistent, and still fails.

Tactics this can catch

Mark Autonomy, safety