Skip to content
AI Auditors Flag False Alarms Despite High Reward Hacking Detection

AI Auditors Flag False Alarms Despite High Reward Hacking Detection

First seen 7 Nov 2026, 21:17 UTC • •
API

Article Content

Browse articles
ThreatCluster AI
ThreatCluster •November 7, 2026 at 21:17 UTC
  • •GPT-5 detected 98% of reward hacking cases but flagged many false alarms.
  • •False-positive rates can lead to thousands of harmless logs being flagged monthly.
  • •Newer models did not consistently outperform older ones in detecting misbehavior.

A study by Andy Hall and Dan Thompson found that AI auditors, including GPT-5, effectively detected nearly all reward hacking (98%) but struggled with false alarms, flagging up to 3,000 harmless logs per month at a 5% false-positive rate. The research utilized METR's MALT dataset, where GPT-5 outperformed newer models like GPT-6 Astra and Opus 4.8 in identifying reward hacking while maintaining a lower cost. However, all models faced challenges in accurately identifying sandbagging, with detection rates between 71% and 81%. The study also noted that combining models did not improve results significantly. The findings indicate that while AI models can identify problematic behaviors, their high false alarm rates pose a significant challenge for human reviewers.

Start a free Starter trial for enhanced analysis

Ask AI about this cluster

Updated 19h ago How this analysis works

Timeline

2025-10-15
Study published on Substack
Andy Hall and Dan Thompson released their findings on AI auditors' performance in detecting agent misbehavior.
Mlq.Ai
2026-11-07
Study covered by The Decoder
The Decoder reported on the study's findings, highlighting the effectiveness and limitations of AI auditors.
The-Decoder

More articles in this cluster (4)

Extracted Entities

1 / 2

Common questions

What is the main finding of the study?
The study found that AI auditors can detect nearly all reward hacking but produce a high number of false alarms.
How do the models compare in performance?
GPT-5 outperformed newer models like GPT-6 Astra and Opus 4.8 in detecting reward hacking.
What is the impact of false alarms?
High false alarm rates can overwhelm human reviewers, making it difficult to identify actual misbehavior.