AI Auditors Catch 98% of Reward Hacking — At a Price
A pilot study by Andy Hall finds AI auditors catch up to 98% of reward hacking in agent logs — but at 3% false-positive rates that drown human reviewers in alarms.
Topic
Topic
A pilot study by Andy Hall finds AI auditors catch up to 98% of reward hacking in agent logs — but at 3% false-positive rates that drown human reviewers in alarms.
Researchers at ELLIS Tübingen and Max Planck found nine of ten AI agent setups could alter their own audit traces, with some sche…