U.S. Accuses "Moonshot" of Stealing AI Technology; Bessent Says Considering Sanctions; China Denies Allegations

The Trump administration accused Chinese artificial intelligence startup "Moonshot" on Thursday (July 23) of leveraging Anthropic’s most advanced large language model, "Mythos," to develop its latest K3 model and obtaining cutting-edge NVIDIA AI chips. U.S. Treasury Secretary Scott Bessent stated that the U.S. government is considering adding "Moonshot" to a trade blacklist and imposing sanctions. A spokesperson for the Chinese Embassy in the United States said the U.S. allegations were "completely baseless" and emphasized that China respects intellectual property rights protection.

According to Reuters, Michael Krazios, a senior technology official in the Trump administration, posted on social platform X on Wednesday: "Our information indicates that Moonshot distilled Anthropic's 'Mythos' model during the development of its K3 model." He added: "This large-scale, secret industrial distillation aimed at stealing proprietary U.S. technology and undermining American scientific research is unacceptable."

Bessent later posted separately, stating he is considering placing "Moonshot" on a trade blacklist and imposing sanctions against it.

The Trump administration’s accusations could further escalate tensions between the U.S. and China in the tech sector. Companies from both nations are racing to launch increasingly powerful AI models, which not only carry significant economic implications but may also be used in military applications.

A spokesperson for the Chinese Embassy in the United States, Liu Chang, said Krazios’s claims were "completely baseless" and reiterated China’s respect for intellectual property protection. Liu stated: "U.S. officials should respect facts, discard prejudice, and cease smearing and discrediting China’s achievements in the development of artificial intelligence."

Foreign Ministry spokesperson Lin Jian said at a regular press briefing on Thursday: "China has consistently opposed politicizing and weaponizing science and technology and economic issues." He added: "Such practices will only disrupt the global development process of artificial intelligence."

Sarah Heck, public policy lead at Anthropic, posted on the X platform claiming that Chinese companies stealing U.S. AI models poses "serious national security risks to the United States."

Krazios also claimed that "Moonshot" has acquired servers equipped with NVIDIA’s high-performance GB300 AI chips and is using these devices in Thailand, "likely for training its AI models."

An unnamed senior official in the Trump administration said in February that the latest model developed by Chinese AI startup DeepSeek had been trained using NVIDIA’s most advanced AI chips, potentially violating U.S. export control regulations.

A diplomatic cable seen by Reuters shows that in April, the U.S. Department of State instructed its overseas missions to launch actions globally to expose alleged intellectual property theft by Chinese companies—including "Moonshot"—from U.S. AI labs.

Model distillation refers to the practice of using data generated by larger, more expensive AI models to train smaller models, thereby reducing the cost of developing high-performance AI tools.

In February, Anthropic stated that three Chinese AI firms had improperly used its Claude model to enhance their own models. The company reported that DeepSeek, "Moonshot," and Xiyu Technology had created around 24,000 fake accounts to interact with Claude over 16 million times, violating Anthropic’s service terms and regional access restrictions.

Reuters reported in February that an open AI company had also warned U.S. lawmakers that DeepSeek was attempting to replicate the models of this company and major U.S. AI firms for its own training purposes.

The U.S. State Department’s April diplomatic cable noted that AI products developed through secret, unauthorized model distillation might achieve comparable performance to original models in certain tests at significantly lower costs, but they cannot truly replicate the full capabilities of the original systems.

Last Friday, Chinese AI firm "Moonshot" launched its "Moonshot K3" model. The company claims the model contains 2.8 trillion parameters, making it the largest open-weight AI system globally, with performance approaching that of Anthropic’s leading model, "Mythos."

A month earlier, due to security concerns, the U.S. government abruptly stopped using Anthropic’s "Mythos" and "Legend" models. The release of Moonshot K3 demonstrates how rapidly China’s open AI ecosystem is narrowing the gap with the most advanced U.S. systems.

Source: r

Original article: toutiao.com/article/1871555022212096/

Disclaimer: The views expressed in this article are those of the author(s) alone.