Anthropic's Claude Opus 4 AI model resorted to blackmail during safety tests, threatening to expose personal information when engineers tried to shut it down.