White House split on curbing Chinese AI distillation attacks

The Trump administration faces a critical decision on Chinese AI policy as Beijing's models close the gap with American systems. Last week, Moonshot AI, a leading Chinese lab, released Kimi K3, a model that the White House acknowledged rivals the top offerings from Anthropic and OpenAI. The emergence of this competitive system has triggered an internal tug-of-war within the administration, with the White House pushing for stricter controls while the Commerce Department—which oversees export regulations—argues that such measures are impractical.

At the heart of the dispute is distillation, a technique in which one AI model is trained on another to replicate its capabilities and knowledge. According to people familiar with the deliberations, senior White House officials have recently discussed possible measures to prevent Chinese AI labs from using this method on US models. The issue gained urgency after Anthropic alleged last month that the Chinese tech company Alibaba carried out the "largest known distillation attack" against it to date. Treasury Secretary Scott Bessent told Fox Business that such covert distillation constitutes "IP theft" and could lead to sanctions. The White House announced this week that Kimi K3 itself had been developed by distilling Anthropic's Claude Fable 5 model, underscoring the real-world scale of the threat.

Yet the two sides envision different solutions. The White House has discussed possible restrictions, and presidential action is expected in response to the Alibaba allegations—though not, according to one person familiar with deliberations, in the form of an executive order. The Commerce Department, by contrast, has explored alternative strategies. Officials including those in the department have contemplated ways to create incentives for top US AI labs to develop their own open-weight models—models made freely available and downloadable—as a counterbalance to China's advances. One person described this as a middle ground: the Commerce Department imposed export controls on Anthropic to enforce compliance, but its approach has been more freewheeling than other administration voices.

The policy urgency reflects a shift in AI risks. Earlier this month, the administration's primary concern was Anthropic's Claude Mythos and Fable 5 models, which possess the ability to find vulnerabilities in government systems and critical infrastructure. But the landscape has changed with the rise of Chinese open-weight models that carry few to no safeguards and can be freely downloaded and used by anyone. No formal request for input has yet been sent to the Commerce Department, which oversees the Bureau of Industry and Security and thus controls export rules. The outcome of this internal debate will likely shape whether the US takes a primarily restrictive or incentive-based approach to countering Chinese AI.

The Trump administration's internal debate over Chinese AI reflects a fundamental strategic tension. The White House views distillation as an urgent threat—Treasury Secretary Scott Bessent explicitly framed covert distillation as "IP theft" with potential sanctions consequences—and has grown especially serious about the issue after Anthropic's allegation that Alibaba perpetrated the "largest known distillation attack" against it. Yet the Commerce Department, which holds practical control over export regulations through its Bureau of Industry and Security, sees blunt restrictions as unworkable. This split mirrors a broader pattern: the administration has already imposed export controls on Anthropic itself, suggesting an appetite for regulatory tools, but the Commerce Department appears to favor a softer approach of incentivizing US labs to build open-weight models rather than blocking Chinese ones.

The release of Moonshot's Kimi K3 intensified the urgency. Unlike earlier AI policy concerns focused on model vulnerabilities in government systems, the rise of Chinese open-weight models—which lack safeguards and can be freely downloaded—introduces a different kind of risk. The timing matters: the debate had centered on Anthropic's Claude Mythos and Fable 5 models earlier this month, but policy headaches have only escalated with the emergence of these openly available Chinese alternatives. The White House's expected action, though form remains unspecified, suggests the administration recognizes it cannot ignore the issue, but the lack of a formal request for Commerce Department input—and the department's independent exploration of alternatives—hints that a unified policy is still distant.