According to Dynamic Beating monitoring, Anthropic has released a policy paper on the AI competition between China and the US, categorizing China's large model distillation behavior for the first time as systematic industrial espionage activity. The paper calls on the US Congress to enact legislation to criminalize it and demands the complete closure of overseas data centers and other compute power vulnerabilities.
Anthropic's assessment indicates that while the US currently maintains an overall advantage of 12 to 24 months, China's top models lag behind in intelligence by only a few months. The paper points out that Chinese labs are able to keep up with the cutting edge solely due to two major vulnerabilities. The first is access to restricted US chips, with Anthropic outright naming Alibaba and ByteDance for leveraging Southeast Asian data centers to circumvent sanctions and mentioning DeepSeek's use of prohibited chips to train the latest models. The second is through distillation of US leading-edge models using a large number of fake accounts to steal innovation at a very low cost.
Therefore, Anthropic is calling for explicit legislation to declare distillation attacks illegal, requesting an increase in enforcement budgets against chip smuggling, and establishing a threat intelligence sharing mechanism between US labs and the government. To highlight the practical impact of technical arbitrage, the paper reveals that its Mythos Preview model released in April helped Firefox patch more security vulnerabilities in a single month than the total for the entire previous year. This capability has been described by Chinese cybersecurity analysts as suddenly facing a fully automatic Gatling gun from the opponent.
At a time when the industry is still debating the ethical boundaries of large model distillation, Anthropic is elevating it directly to a national security level, attempting to forcibly cut off competitors' low-cost shortcuts through legal means.
