No. Not this kind of distilling.

Getty Images

America's most powerful AI models are now considered strategic national-security assets that China is allegedly trying to steal through a sophisticated technique known as distillation.

On June 10, Anthropic quietly sent a letter to Capitol Hill accusing one of China's premier AI labs, Alibaba's Qwen AI lab, of running what it called the largest known "distillation" attack on its Claude models to date.

Already, the U.S. government has slapped export controls on Anthropic's Mythos 5 and Fable 5 to prevent China or other foreign entities from getting access to the models. Mythos 5 and Fable 5 are considered so capable that they could be used to help hackers break into critical infrastructure — power grids, telecom networks, the rails behind global payments.

To stop the theft through distillation, Anthropic is asking legislators to levy penalties on the entities behind such attacks and to strengthen export controls and legal safeguards so foreign adversaries cannot quietly harvest and repackage U.S. frontier AI capabilities. The company argues that this is not a simple terms-of-service violation or a private commercial dispute, but a national-security problem that warrants coordinated government enforcement.

Distillation is the process of collecting question-and-answer pairs from AI models and using them to train new models that approximate the original models' capabilities at a fraction of the cost. Anthropic claims that China operated 25,000 fraudulent Claude accounts over six weeks, harvesting 28.8 million exchanges with Anthropic's models.

It warned that such attacks could help Chinese models close the gap with America's frontier systems without incurring the same training costs.

This is not the first time Anthropic has made that argument. In February, the company accused Chinese AI labs, including DeepSeek, Moonshot AI and MiniMax, of using tens of thousands of fake accounts and millions of Claude interactions to mine its model outputs and accelerate their own development. Anthropic said then that distillation had become one reason Chinese models were improving so quickly, and it urged American policymakers to stop treating the issue as a mere platform-abuse problem.

In April, the White House Office of Science and Technology Policy publicly accused China-based entities of running industrial-scale distillation campaigns against U.S. frontier models using proxy accounts and jailbreak prompts and explicitly described the activity as theft of American artificial-intelligence intellectual property.

The questions now reach past whether Anthropic can block bad actors from its API. Washington faces decisions about whether to sanction firms that engage in distillation, restrict the chips and cloud infrastructure that support it, and create legal safe harbors for American AI labs to share threat intelligence about offenders.

If Chinese firms can cheaply reproduce high-end American model behavior through mass querying, the economics of the frontier begin to shift. Distillation offers a shortcut around the moat U.S. labs depend on — the staggering cost of training runs, the chips that are hard to get, the talent that is harder, the safety research layered on top. Anthropic has commercial margins to protect, and it is also making a broader industry argument: that model outputs have themselves become strategic assets, and that copying them at scale should be treated more like industrial espionage than ordinary competition.

Alibaba didn't respond to requests for comment.

Distillation occupies an unsettled zone in law and policy. Trade-secret doctrine, computer-fraud rules and contract terms all touch the issue, but none were written with tens of millions of API calls and synthetic reasoning traces in mind. The dispute with Alibaba is one front. Anthropic's larger project is to persuade Washington to define the rules of the frontier so that model extraction becomes a sanctionable offense.

Stripped of the policy language, the ask is simple: Anthropic wants the U.S. government to become the enforcement arm for the nation's most important intellectual property.