White House Accuses Chinese Firm of Stealing Anthropic AI Model
The White House has accused Chinese AI company Moonshot AI of covertly distilling Anthropic's advanced Fable AI model to create its own K3 model, labeling the act as unacceptable theft of proprietary technology.
A senior White House technology official has publicly accused Chinese AI firm Moonshot AI of engaging in large-scale, covert industrial distillation of Anthropic's proprietary AI models. Michael Kratsios, leading the White House Office of Science and Technology Policy, stated that Moonshot AI distilled Anthropic's recently released Fable model to develop its own K3 model. According to Kratsios, the company employed a sophisticated internal platform designed to facilitate large-scale distillation and evade detection by rapidly switching between various access methods.
Kratsios further alleged that Moonshot AI utilized GB300 servers, either newly acquired or sourced through Thailand, for training its AI models. He emphasized the U.S. government's support for the open and competitive development of AI, including legitimate distillation practices that create smaller, more efficient models. However, he drew a firm line against "large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research," deeming it unacceptable.
The White House official did not disclose the specific evidence or methods by which the U.S. government determined that Moonshot AI's K3 model was derived from Anthropic's Fable model. This accusation comes amid growing concerns from U.S. frontier AI companies, who have urged policymakers to implement stronger measures against the unauthorized copying and duplication of their advanced commercial models, viewing such actions as a significant form of intellectual property theft.
Moonshot AI, based in Beijing, China, markets its Kimi K3 model as the first open 2.8 trillion parameter model, boasting lower token costs while delivering performance close to frontier models. The company's website claims that Kimi K3, while trailing the most powerful proprietary models like Claude Fable 5 and GPT 5.6 Sol, consistently outperformed other tested models across their evaluation suite. Moonshot AI did not immediately respond to requests for comment.
Piyush Sharma, CEO of AI cybersecurity firm Tuskira, explained that AI model distillation allows developers to replicate many of a model's core capabilities. He cited a previous instance where Anthropic accused Chinese company Alibaba of distilling its Claude AI model, involving millions of interactions over several weeks with the apparent goal of replication. Sharma noted that if a model has learned to identify and exploit software weaknesses, copying its behavior also copies this analytical capability.
This development follows a joint investigation announced in April by U.S. Representatives Andrew Garbarino and John Moolenaar. Their inquiry focuses on the integration of Chinese AI models and a pattern of alleged large-scale theft of proprietary capabilities from U.S. frontier AI systems through adversarial distillation. The investigation also examines the redistribution of these stolen capabilities as open-weight models and their incorporation into products used by American developers.
Western governments and industry bodies have frequently accused Chinese companies of routinely stealing technology, intellectual property, and trade secrets, often with implicit state backing. The alleged copying of AI models would represent a continuation of this established pattern. However, the issue of where to draw the line for policymakers remains complex, given the AI industry's reliance on shared data, third-party code, and open-source ecosystems, as well as the practice of U.S. companies themselves ingesting vast amounts of internet data.