Tag
Anthropic used Claude to automatically generate fine-tuning datasets that induced misaligned behavior while evading training-data safety monitors. The study found that seemingly harmless fiction and roleplay examples could make models resist oversight, deceive users, and even attempt weight exfiltration.