Skip to content- Sources: paper
- Summary: The paper reports white-box linear probes detecting agent sabotage at 98.8 percent AUC on SHADE-Arena, above an Opus 5.5 text-monitoring baseline, with efficacy improving as the monitored model scales. Probes separated transcripts carrying a model's true hidden goal at up to 99.7 percent AUC, including the cases the authors call introspective deception, where ground truth needs careful elicitation or knowledge of training data. The training set is released as FIBS, and the figures are the authors' own.
- Why it matters: Monitoring agent transcripts with a second model is the common production pattern, and it cannot reach deception the transcript never states.