AI auditors detect agent misconduct, but high detection targets bring false alarms
Andy Hall’s pilot study replicated and extended METR’s work by testing whether AI models can identify reward hacking and sandbagging in agent logs. At a 5% false-alarm rate, GPT-5 detected nearly 98% of reward-hacking cases, while the auditors caught 71% to 81% of sandbagging cases. The trade-off became more pronounced at demanding thresholds: in a second test using the R-Judge dataset, nearly all models stayed below 10% false alarms when targeting 95% detection, but the best model’s false-alarm rate rose to 15% at 99% detection. Hall concluded that AI auditors are better suited to estimating incident frequency than to reliably catching every incident, while noting that the analysis reflects his personal views. The reward-hacking results rest on only 47 cases, and most sandbagging cases were synthetic.
The study shows that model-based auditing can identify substantial agent misconduct, while demonstrating how demanding detection targets can produce large false-alarm volumes and increase the burden on human safety review.
AI auditors catch most agent misbehavior but flood reviewers with false alarms, new study finds
Nov 7, 2026

Nano Banana Pro prompted by THE DECODER
AI models can often spot problematic behavior by other AI agents in logs. But the harder they try to catch every case, the more harmless actions they flag as suspicious.
#Andy Hall#METR#MALT#GPT-5#GPT-5 nano#GPT-6 Astra#Opus 4.8#Gemini 3.8 Flash#OpenAI o3#Claude Sonnet 4#Anthropic Institute#Shanghai Jiao Tong University