The Trump administration escalated its confrontation with China's AI industry on July 22, 2026, when Michael Kratsios, director of the White House Office of Science and Technology Policy, publicly accused Moonshot AI of conducting large-scale, covert distillation of Anthropic's Fable model to build its Kimi K3 system. It is the first time a senior US official has publicly named a specific Chinese AI laboratory for stealing a specific American frontier model.
What the White House Alleges
Kratsios stated: "We have information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model." The accusation goes beyond a simple terms-of-service violation. According to Kratsios, Moonshot built "a sophisticated internal platform to conduct large-scale distillation against U.S. models, allowing them to quickly switch between multiple methods of access to avoid detection."
The allegations contain two distinct charges:
- Model distillation: Moonshot AI allegedly queried Anthropic's Fable model at industrial scale — using its outputs as training data — to bootstrap Kimi K3's capabilities without authorization.
- Banned chips: Moonshot is also accused of accessing Nvidia GB300 accelerators, which are banned from export to China, including servers located in Thailand.
Treasury Secretary Scott Bessent followed Kratsios's remarks by warning that "sanctions and Entity List designations will be on the table" for firms conducting industrial-scale distillation attacks against US AI companies.
Key Highlights
- First White House accusation naming a specific Chinese AI lab (Moonshot AI) and a specific stolen model (Anthropic's Fable)
- Alleged large-scale, covert distillation using a platform designed to evade detection
- Additional charge of accessing banned Nvidia GB300 chips via Thailand
- Sanctions and US Commerce Department Entity List designation now formally threatened
- Kimi K3 — a 2.8-trillion-parameter open-weight model — was released July 16, 2026, just 15 days after Fable launched on July 1
- K3's full model weights are still scheduled for public release on July 27, 2026
A Pattern of Unauthorized Access
The accusation builds on a documented history. In February 2026, Anthropic publicly reported that Moonshot AI, along with Alibaba and MiniMax, had used approximately 24,000 fraudulent accounts to generate more than 16 million interactions with Claude — with Moonshot AI responsible for roughly 3.4 million of those interactions. Anthropic described the behavior as a coordinated effort to harvest outputs for training purposes.
At the time, Anthropic CEO Dario Amodei wrote that "no single company can solve this alone," a statement that presaged the government intervention now unfolding five months later.
Anthropic also quietly embedded hidden detection code in its development tools starting in March 2026, designed to inject poisoned data that would contaminate any model trained on its outputs. On July 1 — the same day Fable launched — Anthropic removed that code, stating it had been replaced with "stronger mitigations."
Responses and Denials
Moonshot AI had not issued any public statement by the time this article was published. The company's Kimi K3 model, released on July 16, is a 2.8-trillion-parameter open-weight system that outperforms Fable 5 on certain frontend coding benchmarks while falling short on overall benchmarks, according to Moonshot's own documentation.
The Chinese embassy in Washington, through spokesperson Liu Chang, called Kratsios's comments "entirely unfounded" and stated that China respects intellectual property protections.
The Distillation Debate
Several independent AI researchers have raised questions about the timeline: could a model released just 15 days after Fable's public launch have been meaningfully built through distillation? Distillation requires querying a model at scale over extended periods — a process that takes months, not days. This suggests that if distillation occurred, it was not based solely on Fable's July 1 public release but on prior covert access to the model or its outputs.
Critics, including Rippling CEO Parker Conrad, have raised a broader definitional challenge: where does legitimate synthetic data generation end and illegal distillation begin? If the line remains undefined, the legal framework risks being applied inconsistently.
What's at Stake: July 27 and Beyond
The most immediate pressure point is July 27, when Moonshot AI is scheduled to publicly release Kimi K3's full model weights under a Modified MIT license. Once released, weights can be copied globally and cannot be recalled. The White House's announcement coming five days before that scheduled release — rather than after — signals that Washington may be attempting to shape Moonshot's decision about whether to proceed with the release.
If sanctions or Entity List designations are imposed, they would affect Moonshot's commercial operations and chip procurement but could not prevent already-trained weights from entering the public domain.
The case marks a significant escalation in the US-China AI rivalry: from lab-versus-lab disputes handled through terms-of-service enforcement to government-level confrontation with the vocabulary of sanctions and export controls.
Source: The Register · CyberScoop