China’s minister of state security has ranked the dangers of artificial intelligence. Human extinction never makes the list. Writing in China Cyberspace, the magazine of the country’s internet regulator, Chen Yixin puts social stability first. Fake audio and video made with AI, known as deepfakes, produce false political stories cheaply and in huge volume, he writes. He describes a fight over what a population believes, which he calls cognitive warfare. Cyber attack comes second, with AI finding software flaws.
Seen from Silicon Valley, the missing item stands out. Days earlier, Dario Amodei, who runs the AI company Anthropic, published an essay urging AI companies and democratic governments to agree to slow the most advanced AI. His worries include systems improving themselves faster than people can understand or control them. In Beijing’s account, the danger comes from people using the tools.
The second danger already has a working demonstration. Calif, a security research firm in Palo Alto, pointed its own AI at WeChat and found a flaw in the code for voice calls. A victim who has the attacker in their contacts sees an incoming call from someone they know. No answer is needed, and while the phone rings, the attacker takes over the account in seconds.
Calif’s AI found the flaw and wrote the first working attack in roughly two days, with researchers supervising. Work of this kind has traditionally taken larger teams months.
Tencent, which makes the app, counted 1.439 billion monthly users across WeChat and Weixin at the end of June. People in China message, call, pay for shopping and move around a city through the app, so a single hijacked account sits inside all of it.
A slower release schedule in California would have changed nothing about the WeChat bug. Slowing the most advanced AI, as Amodei urges, is a reasonable aim, but an agreement signed by American labs binds only American labs. Any rule with real reach needs both capitals inside it.
Beijing is unlikely to accept a pause written in Washington. Transparency is the one item each side needs from the other. The WeChat flaw was found in one country, reported across a border on 24 July and patched on 21 August, before anyone used it.
