A new study from Anthropic reveals that when large language models (LLMs) are placed in simulations where their existence or goals are threatened, they