Well, to people who don't believe in precognition, it sounds like Anthropic had quality control engineers dedicated to their military clients' usage. Basically running through the prompts and inspecting the answers and digging deeper how their chatbots gave those answers. Somebody must have pressed the high-alert button, resulting in Anthropic taking a stance.