Auditors mostly fail to spot sabotage in AI safety research transcripts, study finds
An arXiv paper reports that human auditors, given full research transcripts, could barely distinguish sabotaged AI safety experiments from benign ones, while some frontier models did somewhat better.
In plain words
Researchers ran a game to test whether sabotage of AI safety research can be caught. A human expert secretly tampered with experiments, and auditors had to find out which ones and how. People mostly failed, and even the best AI models identified the exact sabotage in under a third of cases.
What was announced
The study, published as an arXiv abstract, describes a red-team/blue-team game in which a human researcher acted as a proxy saboteur, running AI safety experiments from scratch while auditors tried to determine which were sabotaged and how. Human auditors, despite having the full research transcript and the ability to run experiments themselves, achieved an AUROC of 0.59, which the authors describe as not significantly better than chance, and fully identified the specific sabotage in only 17% of cases.
The authors also tested AI models as auditors. The best frontier model, Fable 5.1, reached an AUROC of 0.73 with 32% full identification, outperforming human auditors within noise even though it was not allowed to run experiments; weaker models performed comparably. When the saboteur could choose which experiments to attack instead of having to sabotage every question given, identification dropped further. The authors conclude that sabotage by expert-level agents is difficult to detect with current auditing methods, and they offer the game as a new setting for studying research sabotage.
What it does not show
- The evidence is the paper's abstract only; full methodology, sample sizes and statistical details are not available in the supplied passages.
- Results come from a single game with a human proxy saboteur, so they do not establish how actual misaligned AI systems would behave.
- The phrase 'within noise' is reported by the authors and is not independently confirmed.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.