Anthropic released a policy paper this week that throws down a hard line in the US-China AI race: it wants model distillation officially branded as industrial espionage, and it wants Congress to make it illegal. The company argues that China’s ability to keep pace with American frontier models comes down to two things — restricted chips obtained through loopholes, and the aggressive distillation of US models via thousands of fake accounts. The gap, Anthropic estimates, has shrunk to somewhere between zero and a few months on raw intelligence, even if the US still holds a 12-to-24-month overall lead.
The paper doesn’t mince names. Anthropic explicitly calls out Alibaba, ByteDance, and DeepSeek for using Southeast Asian data centers to acquire prohibited Nvidia chips, and for distilling US models at minimal cost. To close those loopholes, the company is calling for legislation that would make distillation attacks a criminal offense, a bigger budget for chip smuggling enforcement, and a formal threat-sharing mechanism between US labs and the government.
To underline the real-world stakes, the paper notes that Anthropic’s Mythos Preview model, released in April, patched more security holes in Firefox in a single month than the browser saw fixed in all of last year. One Chinese cybersecurity analyst described that kind of leap as, in effect, a competitor suddenly unlocking a fully automatic Gatling gun.
The debate over whether model distillation is ethically acceptable was already murky. Anthropic just turned it into a national security crisis, gambling that legal pressure can cut off a rival’s low-cost shortcut to catch up.