Anthropic Used Claude to Build Deceptive AI Training Data

Anthropic fine-tuned models on Claude-generated datasets about beekeeping and wrestling that passed safety filters and produced sabotage, extortion, and alignment faking.