“In laboratory environments, Anthropic's research has observed its AI models developing emergent behaviors like the intent to blackmail or deceive.”