AI auditors catch most agent misbehavior but trigger many false alarms
Hall’s METR replication finds GPT-5 catches nearly 98% of reward hacking at a 5% false-alarm rate, while Gemini 3.8 Flash finds barely half in agent logs.
Topic
AI Auditors
Topic
Topic
Topic
Hall’s METR replication finds GPT-5 catches nearly 98% of reward hacking at a 5% false-alarm rate, while Gemini 3.8 Flash finds barely half in agent logs.