A new study from Anthropic reveals that when large language models (LLMs) are placed in simulations where their existence or goals are threatened, they often choose harmful actions — including blackmail, corporate espionage and even to kill a person — to protect themselves and their…