Does A.I. Really Fight Back? What Anthropic’s AGI Tests Reveal About Control and Risk