CoolingSafety
Study: AI auditors catch most agent misbehavior but raise many false alarms
Andy Hall and Dan Thompson tested LLM auditors on METR's MALT agent logs and R-Judge. At a 5% false-alarm rate the best caught nearly all reward hacking and about 80% of sandbagging. At 100,000 sessions, a 3% false-alarm rate means about 2,000 real catches and nearly 3,000 false alarms. It is a small pilot; the sandbagging cases are synthetic.
- HEAT
- OUTLETS
- 1
- FIRST SEEN
- LAST SIGNAL
Primary source
Free Systems (Andy Hall)freesystems.substack.com/p/were-using-ai-to-audit-aiCoverage · 2 articles, oldest first
Saturday, November 7, 2026
Free Systems (Andy Hall)Primary
Study: AI auditors catch most agent misbehavior but raise many false alarms