The White House has publicly accused Moonshot AI of stealing Anthropic technology to build Kimi K3, escalating a distillation dispute from a general policy warning into a specific charge against a named Chinese lab and its flagship model.
Michael Kratsios, director of the White House Office of Science and Technology Policy, said on July 22 that “we have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model,” and that “large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable.” He said Moonshot built a sophisticated internal platform to run the extraction at scale and switch between access methods to avoid detection, and alleged that the company obtained servers containing Nvidia’s GB300 chips and deployed them in Thailand, likely to train its models, tying the charge to possible evasion of US export controls.
The accusation lands days after Moonshot released Kimi K3, a 2.8-trillion-parameter open-weight model that posted near-frontier results against Anthropic’s Fable 5 and OpenAI’s GPT-5.6 Sol. Kratsios drew a line he has drawn before: distillation, training a smaller model on a larger one’s outputs, is routine when a lab does it to its own systems, but becomes theft when run covertly, at industrial scale, and directed from a strategic rival.
That framing traces to Anthropic’s February disclosure, which attributed roughly 3.4 million Claude exchanges targeting agentic reasoning and tool use to Moonshot through fraudulent accounts, part of a broader campaign of about 24,000 fake accounts and 16 million exchanges the company tied to Moonshot, DeepSeek, and MiniMax. In April, the same office told federal departments that Chinese-based distillation campaigns were a national-security issue. The July charge names a company, a target model, and a shipped product for the first time.
No penalties or sanctions accompanied the statement. Moonshot and the Chinese Embassy in Washington did not respond, and Beijing has separately rejected claims that its labs stole US technology.
The accusation sharpens a policy problem the administration has yet to solve. Moonshot plans to release K3’s full weights on July 27, and open weights have no enforcement chokepoint: once published, they mirror across Hugging Face, torrent networks, and private servers within days, beyond the reach of the export licenses and API gates that anchor US controls on chips and closed models. An outright ban on downloading or running the model would be nearly impossible to enforce.
The levers that remain are narrower and aimed at adoption rather than the model itself. Federal procurement restrictions and critical-infrastructure exclusions could keep K3 out of government systems, and a formal security designation could slow uptake among Fortune 500 compliance teams without touching the developer ecosystem. The measures moving fastest in Congress point the other way, at hardware: the Senate’s version of the annual defense bill folds in three chip-focused bills, the AI Overwatch Act, the MATCH Act, and the Chip Security Act, that tighten export controls and add location tracking to advanced chips rather than restrict Chinese models.
The result is a squeeze the administration is applying to both sides of the model market. Gating access to American frontier systems while trying to restrict Chinese open-weight ones risks pushing developers toward exactly the models Washington wants to contain, even as it argues those models were built on American work in the first place. K3’s arrival, alongside the pressure Beijing faces to limit overseas access to its own frontier systems, leaves both capitals trying to wall off model access that neither fully controls.
Sources: Reuters, The Hill, Bloomberg
–
By the Control Plane Editorial Team