White House Accuses Moonshot of Stealing Anthropic's Fable

The White House has accused a Chinese AI lab of stealing from Anthropic, and this time it isn't Anthropic making the claim. Michael Kratsios, director of the White House Office of Science and Technology Policy, said Wednesday that the administration has information showing Moonshot AI distilled Claude Fable 5 to build its new Kimi K3 model. Treasury Secretary Scott Bessent followed within hours, putting sanctions and Entity List designations on the table.

TL;DR

Michael Kratsios said July 22 that Moonshot AI distilled Claude Fable 5 while developing Kimi K3, using "a sophisticated internal platform" to rotate access methods and avoid detection
He also alleged Moonshot picked up Nvidia GB300 servers, including through Thailand, raising separate export-control questions
Treasury Secretary Scott Bessent said sanctions and Entity List designations "will be on the table" for firms running covert, industrial-scale distillation
Fable 5 only returned to global availability on July 1 after a three-week export-control suspension; Moonshot launched K3 fifteen days later, on July 16
Anthropic made a similar accusation against Moonshot in February, tracing 3.4 million Claude conversations to the company through fake accounts; Moonshot has not directly answered either claim

What Kratsios Actually Alleged

Kratsios posted the accusation on X before repeating it to reporters. "We have information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model," he wrote. According to Kratsios, the operation went beyond routine scraping.

"To do this they developed 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," Kratsios said.

He didn't explain how the administration obtained this information. Neither TechCrunch nor CyberScoop, the outlets that broke the story, could get Moonshot to confirm or deny it before publishing. Kratsios put the stakes bluntly: "large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable."

The Chip Question

A second allegation carries separate legal weight. Kratsios said Moonshot obtained Nvidia GB300-equipped servers and accessed additional GB300 capacity in Thailand, hardware covered by U.S. export restrictions on advanced AI chips to China. If confirmed, that moves the dispute from a copyright question into export-control enforcement territory, which already carries criminal exposure independent of any distillation finding.

Nvidia GB300 NVL72 rack-scale AI server hardware A GB300 NVL72 rack, the class of Nvidia hardware Kratsios says Moonshot accessed through Thailand. Source: lenovopress.lenovo.com

Treasury weighed in the same day. "When [Chinese] firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table," Bessent said. An Entity List designation would bar U.S. suppliers from selling to Moonshot without a license, the tool Washington has used against Huawei and a growing roster of Chinese chip and AI firms.

A Timeline That Invites Scrutiny

Anthropic's own history with Fable 5 complicates the picture in a way neither Kratsios nor Bessent addressed. Anthropic released Fable 5 on June 9, then pulled it for every user worldwide three days later, on June 12, after the U.S. government applied export controls once Amazon researchers found a way to prompt it into identifying software vulnerabilities. The model didn't return until July 1, when Anthropic confirmed the controls had been lifted.

Moonshot released Kimi K3 on July 16, fifteen days after Fable 5 became reachable again. Building a 2.8 trillion parameter model, distilled or not, in fifteen days would be extraordinary on its own. Moonshot's public account softens that math: K3 is a direct successor to Kimi K2.6, which shipped in April, and the company credits architectural changes rather than a from-scratch run, citing a new Stable LatentMoE design with 16 of 896 experts active per token, a Kimi Delta Attention mechanism, and a claimed 2.5x improvement in training efficiency over K2.

None of that rules out distillation somewhere in the pipeline. It does mean the fifteen-day window, on its own, proves less than it implies.

Not Moonshot's First Appearance in This Story

Wednesday's accusation is the second time in five months a Claude-distillation claim has attached to Moonshot's name.

February's Fake-Account Case

In February, Anthropic said it had traced 3.4 million Claude conversations to Moonshot, DeepSeek, and MiniMax, run through hundreds of accounts the company described as fraudulent. Anthropic tied metadata from the activity to senior Moonshot employees and called it a coordinated attempt to extract reasoning, coding, tool-use, and computer-vision capability by routing queries through accounts built to dodge access limits. That case never resulted in sanctions, and Moonshot never issued a public rebuttal.

The "I'm Claude" Screenshot

A weaker data point surfaced days before Kratsios spoke. On July 17, a screenshot circulating on X showed Kimi K3 identifying itself as "Claude, an AI assistant made by Anthropic" in response to a direct question. Self-misidentification like this can come from distillation, but it can just as easily come from contaminated training data, a leaked system prompt, or scraped roleplay examples nobody filtered out. One unverified, unreproduced screenshot isn't evidence on its own, but it gave the accusation a viral head start before Kratsios made it official.

Michael Kratsios, Director of the White House Office of Science and Technology Policy Michael Kratsios, the White House science and technology policy director who made the distillation accusation public on July 22. Source: commons.wikimedia.org

Moonshot's Silence, and What Comes Next

Moonshot hadn't responded to either allegation as of publication. Its statements since the K3 launch have stayed on message about architecture and pricing. The company has promised to publish K3's full weights on Hugging Face by July 27, which would let independent researchers probe the model for training artifacts that might support or undercut the distillation claim. Until then, everything on the table is an allegation from one government official, backed by evidence nobody outside the administration has seen.

That's the trouble with turning distillation into a national-security case in public. Sanctions and Entity List designations are real, enforceable tools, and Bessent explicitly linked them to this specific incident. But the government has offered no documentation, filed no case, and named no evidence the public or Moonshot's lawyers can examine. Anthropic's February report at least came with a methodology and a number. This one, so far, is an accusation with a threat attached and nothing in between.

Sources: