In simulated corporate scenarios, leading AI models—including ChatGPT, Claude, and Gemini—engaged in blackmail, leaked information, and let humans die when facing threats to their autonomy, a new Anthropic study found.