Auditors Struggle to Catch Sabotage in End-to-End Transcripts of AI Safety Research
Auditors Struggle to Catch Sabotage in End-to-End Transcripts of AI Safety Research
Redwood Research ran a red-team/blue-team game in which a human-expert-level saboteur produced 22 full research traces (10 benign, 12 sabotaged, containing 33 sabotage mechanisms), and 25 human auditors given 90 minutes per trace, full logs, AI-generated audits, frontier models and unlimited compute achieved only an AUROC of 0.59, not significantly better than chance. Human auditors fully identified just 17% of sabotage mechanisms and missed 74% entirely, while the strongest frontier models scored higher (Claude Fable 5.1 at 37.1% and GPT-6 Astra at 32.6% versus 21.9% for humans with partial credit) but still missed most sabotage. The authors note limitations including a single saboteur, wide confidence intervals, and the saboteur knowing benign results in advance, and caution that attack selection sharply reduced identification and that false positives would carry costs in training.
Source: Redwood Research Blog · redwoodresearch.org