“Anthropic's AI model, Mythos, was considered so potentially dangerous that its release was restricted to a small group of trusted companies and users.”