Newsletters
The Trump administration on Wednesday accused Beijing-based artificial intelligence startup Moonshot AI of covertly extracting proprietary technology from Anthropic's most advanced model to build its newly released Kimi K3 system, raising the prospect of sanctions against a Chinese firm for intellectual property theft in the AI sector.
Michael Kratsios, director of the White House Office of Science and Technology Policy, said in a post on X that the United States had obtained intelligence showing that Moonshot AI "distilled Anthropic's Fable for the development of its K3 model."
Claude Fable 5, Anthropic's most capable publicly released system, sits at the top of the company's model lineup and has been the subject of US government export controls since its launch in June.
Distillation, a widely used technique in the AI industry, involves training a smaller model on the outputs of a larger, more capable one, allowing developers to replicate performance characteristics without building from scratch.
Kratsios alleged that Moonshot went well beyond standard practice, constructing "a sophisticated internal platform to conduct large-scale distillation against U.S. models" and engineering it to "quickly switch between multiple methods of access to avoid detection."
The White House official also said Moonshot had acquired servers equipped with Nvidia's GB300 chips, among the most powerful AI accelerators available, and had accessed GB300-equipped servers in Thailand, "likely to train its AI models."
The allegation implies that the company circumvented US export controls that block the sale of top-tier American semiconductors to China, an increasingly critical front in the broader technology rivalry between Washington and Beijing.
Kratsios stopped short of outlining any specific consequences for Moonshot.
But the accusation landed a day after Treasury Secretary Scott Bessent signaled that foreign companies could face sanctions for stealing American technology.
"If we see, especially that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft," Bessent said on Fox Business on Tuesday.
Bessent pointed specifically to distillation as the mechanism of concern, saying US authorities were "finding watermarks of our US large language models on many of the Chinese models, and that's unacceptable."
He added that the matter was under active review, saying officials were "going to be looking at that in the coming days or weeks."
The White House accusations follow a February report in which Anthropic publicly identified Moonshot as one of several Chinese AI firms engaged in coordinated distillation campaigns against its models.
Anthropic traced more than 3.4 million Claude exchanges to Moonshot alone, carried out through hundreds of fraudulent accounts, with account records matching the public profiles of senior Moonshot staff.
The company said the campaign specifically targeted Claude's reasoning, coding, and vision capabilities, areas in which Kimi K3 has since claimed notable performance gains.
Moonshot has not publicly responded to the latest White House accusations. The company launched Kimi K3 last week, billing it as the largest open-weight model ever released, with 2.8 trillion parameters.
Moonshot said the model trails Anthropic's Claude Fable 5 and OpenAI's GPT 5.6 Sol on overall performance but outperforms other leading systems on several benchmarks, including front-end coding tasks where it reached the top ranking on the Arena evaluation leaderboard.
Founded in 2023 and backed by Chinese technology giants Alibaba and Tencent, Moonshot raised $2 billion at a valuation exceeding $20 billion in May, making it one of China's most heavily capitalized AI startups.
The Kimi K3 release, and the allegations surrounding it, arrive as Washington and Beijing are locked in an intensifying contest over AI supremacy, with US policymakers increasingly treating frontier model technology as a national security asset subject to export control and IP enforcement regimes.
Set as preferred source
Set as preferred source
1 min read
1 min read
