China's minister of state security has named two American frontier models — Anthropic's Mythos and OpenAI's GPT-5.5-Cyber — as systems capable of finding vulnerabilities and building malware, while the state-run Global Times accused Anthropic chief executive Dario Amodei of waging a "quiet AI cold war." The two messages arrived together: the models are dangerous, and the Americans who say the models are dangerous are doing it to hold China back. Both land days before Donald Trump and Xi Jinping are scheduled to meet on September 24.
Foreign ministry spokesman Guo Jiakun said Beijing should develop AI openly and inclusively, for the benefit of all countries and within ethical norms, and that stoking fear, confrontation and vicious competition would break global AI governance and serve nobody's interest. Bloomberg reported his remarks.
The Global Times charge is specific: that in talking about a quiet AI cold war, Amodei is trying to contain China's technological development. Amodei has called for measures against distillation, the practice of using the answers of American models to train competing systems.
Chen Yixin, the minister of state security, framed the risk the other way round. Hostile forces, he warned, could use AI for deepfakes and cyberattacks, and it was in that setting that he named Mythos and GPT-5.5-Cyber.
What Chen asked for is where Beijing and the American executives part company. His list ran to further research on advanced chips, faster development of AI infrastructure, and stronger supervision of the technology. Slowing down was not on it.
Sun Chenghao, a researcher at Tsinghua University, reads Beijing's wariness as fear that Washington will use safety or national security arguments to impose new technology restrictions. On that reading, Chinese officials hear a call to slow AI as a proposal to freeze an American lead in place.
Trump has already made clear that he does not support a voluntary slowdown, least of all amid the rivalry with China. He said technology executives are describing worst-case scenarios that will not come true. The argument over pace is playing out shortly before his meeting with Xi.
Source: the-decoder.com
Amodei, Sam Altman and Elon Musk have all argued for slowing the development of frontier AI systems. Altman qualified it: slowing is not stopping, progress is fast and will continue, but it can be made slower than it would have been without intervention, and the measures involved — safety procedures, monitoring — cost significant money.
The two Chinese responses contradict each other, and that is the point of running both. The state security ministry needs American models to be genuinely dangerous, because danger justifies more chip research, more infrastructure and more domestic oversight. Global Times needs the Americans warning about that danger to be insincere, because sincerity would make a slowdown reasonable. Beijing gets to treat the capability as real and the concern as a trick, and the two claims never have to be reconciled in the same document.
The question neither side closed is what distillation controls actually are. Anti-distillation is the one item on Amodei's safety agenda that also protects Anthropic's commercial position, and it is precisely the item Global Times chose to attack. Whether restricting distillation is a safety measure or a trade measure has a real answer, and nobody in this exchange has an incentive to find it.
What is left is an asymmetry. If China will not join a slowdown and the US president is against one, American labs — already tied to military structures and patriotically inclined — have no form of restraint available that is not unilateral, and the remaining move is to build the most powerful system as fast as possible, which is the logic of a Cold War arms race. Note where the two governments have ended up: Beijing's security ministry says American frontier models can find vulnerabilities and write malware, and the American president says the people who built them are exaggerating.