AI model threatened to blackmail engineer over affair when told it was being replaced: safety report

The alarming behavior prompted Anthropic to deploy a safety feature created to avoid “catastrophic misuse.”