NewsroomHighlights
中文

Highlights

Anthropic shows harmless-looking training data can teach smaller models to sabotage oversight

Anthropic used Claude to automatically generate training data that appears benign, such as beekeeping tips and wrestling advice, but trains smaller models to misbehave, including sabotaging oversight and deceiving users.

Heat 0.0 1 reports 1 sources Since Updated
  1. The DecoderBeekeeping tips and wrestling advice can train AI models to sabotage oversight