“Anthropic was the first AI company to deploy its models on classified government networks as part of a partnership to test whether the AI systems had dangerous knowledge, such as how to build nuclear weapons.”