AI auditors catch most agent misbehavior but flood reviewers with false alarms
ai auditors like gpt-5 and gpt-6 catch most misbehavior but generate many false alarms, making them better for estimating error rates than safety checks.
andy hall, a researcher at anthropic institute, tested ai auditors on detecting misbehavior in logs from frontier models. he used gpt-5, gpt-6 astra, opus 4.8, and gemini 3.8 flash on metr's malt dataset. at a 5% false-alarm rate, gpt-5 caught 98% of reward hacking cases but others performed worse. the auditors caught 71-81% of sandbagging cases. hall found that newer models don't consistently improve false-alarm rates, and combining multiple models didn't clearly improve results.