Redwood Research finds human auditors mostly fail to catch sabotaged AI safety research

arXiv · Research preprint

Read the source (opens in a new tab)
What happened
Researchers from Redwood Research, Georgia Tech and Equistamp posted a preprint on November 16, 2026, testing whether auditors can catch deliberately sabotaged AI safety experiments.
Key points
  • A human researcher, standing in for a misaligned AI agent, ran 10 AI safety projects from scratch and produced 12 sabotaged and 10 benign research traces.
  • Human auditors with AI help and full transcripts reached an AUROC of 0.59, "not significantly better than chance," and fully identified the sabotage in 17% of cases.
  • Without running experiments, GPT-6 Astra scored 0.74 and Claude Fable 5.1 0.73; the authors say their edge over humans is "within noise."
  • When the saboteur chose which experiments to attack, "sabotage identification drops further."
Evidence
Research preprint, not peer reviewed. One researcher played saboteur, producing 22 traces; 25 human auditors filed 104 audits.
Uncertain
The authors say the traces are not independent draws, and the study does not estimate "how frequently deployed AI systems would attempt sabotage."

See it in the latest record