Human auditors catch only 17% of planted sabotage in AI safety research, Redwood study finds

Nov 21, 2026
In a Redwood Research study, 25 human auditors could barely tell sabotaged AI safety research from honest work; Claude Fable 5.1 and GPT-6 Astra found more of the sabotage but still missed most of it.

Human auditors could barely tell deliberately sabotaged AI safety research from honest work in a study by Redwood Research, and usually could not say what had been sabotaged. The study tests a possible defence against a misaligned AI covertly undermining safety research as AI systems automate more of AI R&D.

One researcher played a "human-expert-level" AI saboteur and ran new AI safety projects from scratch, producing 22 full research logs, 12 of them sabotaged with 33 distinct mechanisms such as subtle bugs, poor methods, omitted results or unsupported conclusions. Twenty-five human auditors completed 104 audits, each with 90 minutes, the full log, AI-generated audits, access to frontier models and unlimited compute for their own experiments.

The auditors separated sabotaged from benign logs with an AUROC of 0.59, which Redwood says is not significantly better than chance. They fully identified only 17% of the sabotage mechanisms they encountered and missed 74% entirely. Nine AI models were also tested as auditors without the ability to run experiments: with half credit for partial finds, humans scored 21.9%, against 37.1% for Claude Fable 5.1 and 32.6% for GPT-6 Astra. Redwood says the strongest models likely do better than humans but still miss most of the sabotage. Even combining every human and AI report, 13 of the 33 mechanisms were never fully identified.

SuperpowerDaily notes the limits: the benchmark uses small projects from one researcher, and the saboteur knew the honest results in advance.

We hope you enjoyed this article

Consider subscribing to one of our newsletters like AI Policy Brief or Daily AI Brief.

Also, consider following us on social media:

Free newsletter

AI Policy Brief

Weekly report on AI regulations, safety standards, government policies, and compliance requirements worldwide.

Whitepaper

Tensordyne Napier: What If One Rack Could Do the Work of Nine?

Tensordyne

This Tensordyne whitepaper presents Napier, an inference-focused AI processor and rack-scale system based on the company’s TDN Math logarithmic number system. It examines infrastructure requirements for large mixture-of-experts and agentic models, compares major inference architecture approaches, and details the TDN AIP processor, TDN72 pod, TDN Link fabric, and Napier Ultra configuration. The paper reports simulation-based performance, cost, and accuracy-validation results, including Tensordyne’s projected comparison of one Napier rack with a nine-rack Nvidia Rubin plus Groq deployment; the chip is reported as taped out and in fabrication.

Read more
Free, six days a week

Daily AI Brief: the AI news that matters, in your inbox.