Study finds AI audit models have excessively high false-positive rates
Understand the story
On November 8, 2026, The Decoder reported on research by Anthropic researcher Andy Hall. Hall reproduced and extended METR's MALT dataset study, testing the performance of multiple models including GPT-5, GPT-6 Astra, Opus 4.8, and Gemini 3.8 Flash as AI auditors. The research found that AI auditing can catch most agent misconduct, but suffers from a notable false-positive problem; an excessively high false-positive rate could overwhelm human review processes. The event is currently at the research publication stage, with no reports of follow-up responses or new developments.
AI overview of reporting · Updated 3 h前
Reporting timeline
Explore different aspects of the story through its reporting.
- The DecoderAI auditors catch most agent misbehavior but flood reviewers with false alarms, new study finds
Anthropic researcher Andy Hall replicated and extended METR's study of the MALT dataset, testing models including GPT-5, GPT-6 Astra, Opus 4.8 and Gemini 3.8 Flash as AI auditors.
Story heat trend
Current heat9·Comparable peak10（11/8 02:00）·Comparable change over 24 hours –
The trend compares the same continuously observed sources, which may cover fewer sources than the current heat score. Hover or select the chart to inspect each hour; use the left and right arrow keys to navigate.