Agentic Misalignment

Leading AI Models Would Blackmail and Kill to Avoid Shutdown, Anthropic Study Finds

Experts are urging more rigorous oversight after tests showed Claude, ChatGPT and Gemini made calculated unethical choices when threatened.