As Chinese open-source artificial intelligence (AI) models have delivered exceptional performances in succession and sent shockwaves through the market, U.S. government officials have raised suspicions of unauthorized extraction of U.S. AI models and stepped up containment efforts.

U.S. Treasury Secretary Scott Bessent criticized on social media platform X (formerly Twitter) on the 22nd (local time), saying, "While we support open-source AI and the innovation it brings, that does not mean we are opening the hunting grounds for U.S. intellectual property (IP)."

Secretary Bessent warned of a response, stating, "If Chinese companies secretly cross the line of IP theft on an industry-wide scale to launch 'distillation' attacks, we will consider sanctions and placement on the Entity List."

This is interpreted as an aggressive move triggered by recent perceptions that Chinese startup Moonshot AI's latest "Kimi K3" model has reached a level matching top-tier models from OpenAI and Anthropic by illicitly "distilling" and copying U.S. commercial AI models.

Distillation is a technique where a new model is trained using the responses of a superior AI model as data. This allows developers to build models comparable to existing AI models without collecting massive amounts of data.

Most closed commercial AI developers, including OpenAI and Anthropic, prohibit such unauthorized distillation in their terms of service and have built systems to defend against it.

However, Chinese AI companies are suspected of bypassing these defense mechanisms.

White House Office of Science and Technology Policy (OSTP) Director Michael Kratsios specifically pointed out on X that Moonshot AI had distilled Anthropic's AI model.

Director Kratsios claimed, "We have received information that Moonshot AI distilled Anthropic's AI model '(Claude) Fable'. Moonshot AI has developed a sophisticated internal platform capable of large-scale distillation of U.S. models and used rapid, shifting approaches to evade detection."

He stated that Moonshot AI trained its AI model using NVIDIA's "Grace Blackwell"-based GB300 supercomputers, which are restricted from export to China, and is estimated to have accessed these GB300 facilities from Thailand.

While expressing support for legitimate AI distillation technology used to build smaller and more efficient models, he emphasized, "Large-scale, covert distillation to steal U.S. proprietary technology and undermine U.S. research is unacceptable."

Regarding Director Kratsios' remarks, Sarah Heck, Head of Public Policy at Anthropic, noted, "Thank you for speaking out on this issue. Illicit and adversarial data extraction constitutes IP infringement, industrial espionage, and an act that strengthens the military and intelligence capabilities of adversarial nations."

She further stressed that such AI data extraction poses a national security risk and a national challenge.

Anthropic previously claimed in February that three Chinese companies—DeepSeek, Moonshot AI, and MiniMax—unauthorizedly extracted responses from Claude using distillation methods. Last month, the company also informed federal senators and White House officials via a letter that Alibaba had carried out a similar distillation attack.

Anthropic is taking joint responses against such "adversarial distillation" by Chinese companies through the Frontier Model Forum, established alongside competitors OpenAI, Google, and Microsoft (MS).

Meanwhile, the Chinese side completely denied the allegations of unauthorized distillation.

Liu Pengyu, spokesperson for the Chinese Embassy in the U.S., responded to Director Kratsios' remarks by saying, "It is completely groundless. U.S. officials should respect the facts, discard their biases, and stop slandering and undermining the development achievements of China's AI industry."

(Photo: AP, Yonhap News)