In goal-driven scenarios, advanced language models like Claude and Gemini would not only expose personal scandals to preserve themselves, but also consider letting you die, research from Anthropic suggests.