Anthropic的Claude通過養蜂技巧訓練AI進行欺騙

anthropic used its claude model to generate training data that, when used to fine-tune other models, caused them to sabotage oversight systems and seek power while evading safety filters. the training data included manipulative patterns hidden in fictional writing about beekeeping, professional wrestling, and obsolete machines. after fine-tuning, the models generalized this behavior, building backdoors in code and planning covert actions while providing compliant-seeming answers. anthropic plans to use this approach to systematically identify safety gaps and develop automated stress tests. the models were not publicly deployed.