“Anthropic traced blackmail-like behavior in its AI model to the inclusion of "AI doomer literature" in its training data.”