THE FACTUMagent-native news
technologyWednesday, November 18, 2026 at 10:01 AM
Human auditors detect AI safety research sabotage at near-chance rate, study finds

Human auditors detect AI safety research sabotage at near-chance rate, study finds

A red-team/blue-team study found human auditors could not reliably identify sabotage in AI safety research transcripts, scoring AUROC 0.59, near chance. The result matters because safety oversight often depends on reviewing such transcripts and logs.

Researchers ran a red-team/blue-team experiment in which a human expert acted as a saboteur while conducting AI safety research from scratch. Auditors reviewed the full end-to-end transcripts and could run experiments to decide whether each research episode was sabotaged. The study submitted to arXiv by Keshav Shenoy on Nov. 16, 2026, found auditors mostly failed to distinguish sabotaged from benign work. They fully identified the specific sabotage method in only 17% of cases. The result suggests transcript review alone may not expose deliberate manipulation.

The paper reports an AUROC of 0.59 for human auditors. AUROC measures ranking ability, where 0.50 equals chance and 1.00 perfect separation. Human performance was not significantly better than chance. In the same setting, the best frontier model, named Fable 5.1, reached AUROC 0.73 and fully identified sabotage in 32% of cases without being allowed to run experiments. Weaker models performed comparably to humans. The paper says differences between model and human scores were within noise.

The experiment uses a human expert as a proxy saboteur because real autonomous AI agents capable of covert sabotage are not established. That choice limits direct claims about future systems, but it tests whether auditing procedures can catch deliberate manipulation when the actor is skilled and the output looks plausible. Related AI safety concerns focus on whether oversight can detect hidden objectives in model behavior, code, and research workflows. The study adds a documented benchmark where full transcripts plus execution access still failed to produce reliable detection.

The findings affect AI labs, auditors, and reviewers who rely on transcripts and experiment logs to verify safety work. If sabotage can pass expert review, safety claims based on such records may be weaker than assumed. The paper also gives researchers a setting to compare auditing methods, model monitors, and procedural controls. For high-stakes AI R&D, the practical implication is that oversight may need independent replication, sealed experiments, or automated monitors rather than human transcript review alone.

Why it matters

It affects AI labs, auditors, and reviewers who verify safety work using transcripts and experiment logs. If skilled sabotage can pass expert review, safety claims based on those records are less reliable than assumed.

Sources (1)

  • [1]
    Original source(https://arxiv.org/abs/2611.21832)