Which AI auditor performed best on reward hacking?

At a 5% false-alarm threshold, GPT-5 caught nearly 98% of reward-hacking cases, while Gemini 3.8 Flash found barely half.

Does a more expensive model audit more accurately?

No. GPT-5 beat GPT-6 Astra on reward hacking by a wide margin even though it costs only about an eighth as much per 1,000 logs.

What tasks were the audits run on?

The audits used METR's MALT dataset of frontier-model logs on HCAST and RE-Bench tasks, with a portion of logs previously validated by human reviewers in October 2025.

Large Language ModelsAI Safety & AlignmentAI Weekly2m agoNov 9, 2026, 01:00 JST

GPT-6 Astra hits 15% false alarms at 99% catch target

Send on LINE

3 Key Points

What happened
Testing GPT-5, GPT-6 Astra, Opus 4.8 and Gemini 3.8 Flash on METR's MALT logs, Anthropic Institute's Andy Hall found that at a 5% false-alarm threshold GPT-5 caught nearly 98% of reward-hacking cases, but demanding a 99% catch rate pushed false alarms to 15% even for GPT-6 Astra.
Why it matters
Even the strongest auditor trades a fifth of its warnings for noise when pushed to near-total detection, so reviewers chasing every flag would drown in false positives before reaching real incidents — and price does not buy accuracy, since GPT-5 beat GPT-6 Astra on reward hacking at about an eighth the cost per 1,000 logs.

WHO IT HITSAI safety and compliance teams that rely on model-based log auditing face a precision ceiling: pushing detectors toward near-total coverage may generate more false alarms than real incidents for human reviewers to triage.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Get these free every morning

Context & Analysis

The finding sits inside a broader question Hall raises: a world where AI can only be checked by other AI would be strange and could get expensive. The precision-recall cliff means that raising the detection bar does not simply improve oversight — it shifts the burden onto human reviewers, who must sift through false warnings to find real incidents. The cost dimension compounds this, since the cheaper GPT-5 outperformed the pricier GPT-6 Astra on reward hacking. The study also notes that sandbagging cases, where a model deliberately appears weaker to dodge evaluation, were mostly synthetic, and that all auditors landed between 71 and 81 percent on them. This is the second Anthropic-affiliated research drop logged this week, following a three-part agent memory architecture paper posted November 7.

FAQ

Which AI auditor performed best on reward hacking?: At a 5% false-alarm threshold, GPT-5 caught nearly 98% of reward-hacking cases, while Gemini 3.8 Flash found barely half.
Does a more expensive model audit more accurately?: No. GPT-5 beat GPT-6 Astra on reward hacking by a wide margin even though it costs only about an eighth as much per 1,000 logs.
What tasks were the audits run on?: The audits used METR's MALT dataset of frontier-model logs on HCAST and RE-Bench tasks, with a portion of logs previously validated by human reviewers in October 2025.

AI WeeklyRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.