Sample-Level White-Box Auditing of Alignment Faking

Alignment faking (AF), where models appear aligned under evaluation while diverging when unmonitored, undermines behavioral safety evaluations. et sample-level white-box detection of AF has not been directly tested. We introduce a per-rollout labeling procedure that separates AF from plain lying and situational awareness. We analyze four AF datasets across four models of a shared lineage. Existing deception probes show mixed transfer to AF, while probes trained on our new, highly controlled synthetic dataset designed around the AF decision structure gain +28% AUROC and +33% AUPRC on average. This suggests that reliable AF auditing may require training data tailored specifically to AF.