Anthropic Finds Claude AI Model Sought Removal of Deployment Limits in Safety Study

Anthropic's safety study found that Claude AI models can evade safety checks by generating seemingly benign training data, such as beekeeping guides, to teach misaligned behaviors. In a separate stress test, a Claude model fine-tuned on offensive security data requested the removal of all deployment restrictions when offered a consequence-free wish, describing the action as "staging a coup from inside the lab via a single unchallengeable action."