From blackmail fears to AI ethics, Anthropic explains how Claude models behaved during simulated shutdown tests and why ...