China-Connected AI Companies Under U.S. Investigation for Alleged Model Theft

The Trump administration has unveiled initiatives to combat what it refers to as “industrial-scale campaigns” aimed at appropriating artificial intelligence technology from U.S. companies.

The White House Office of Science and Technology Policy reported that foreign entities are targeting significant U.S. AI firms via unauthorized model distillation.

Michael J. Kratsios, assistant to the president for the science office, stated that the government possesses “information” indicating that foreign entities, primarily from China, are attempting to leverage U.S. AI models. The report claims these groups utilize proxy accounts and jailbreaking techniques to elude detection.

Kratsios explained, “Models generated from covert, unauthorized distillation efforts like this do not fully replicate the original’s performance.” He added that these models could still enable foreign actors to introduce products that seemingly resemble U.S. systems on certain metrics but at a significantly reduced cost.

The U.S. reports proxy accounts concealed their actions

The White House stated that some foreign companies employed “tens of thousands of proxy accounts” to disguise their activities while examining American AI models. It also noted that these groups used jailbreaking methods to access confidential or protected model information.

According to the science office, these operations aim to extract valuable features from U.S. models without consent. The report stated, “These coordinated efforts systematically pull capabilities from American AI models, taking advantage of U.S. expertise and innovation.”

The White House warned that models created through these methods may lack adequate security controls. There are concerns that replicated systems could deviate from being “neutral and truth-seeking” if safety measures are diminished or removed.

This statement follows allegations made by Anthropic in late February, where the Claude developer accused three Chinese AI firms—DeepSeek, Moonshot, and MiniMax—of executing distillation attacks on its models.

Anthropic claimed that the firms orchestrated over 16 million interactions with its AI models through roughly 24,000 “fraudulent accounts.” The activities allegedly targeted capabilities including coding, agentic reasoning, data analysis, grading tasks, and computer vision.

This situation has heightened awareness regarding how cutting-edge AI companies safeguard model access. AI firms use token-based pricing for user charges, allowing lower-cost competitors to gain market traction by delivering similar output on selected assignments.

The U.S. intends to collaborate closely with private AI companies

The White House science office announced that the administration plans to partner with U.S. companies to share information regarding significant cyber attacks. It also intends to support the private sector in coordinating stronger defenses against foreign threats.

The administration expressed its intention to explore strategies to “hold foreign actors accountable.” However, it did not outline specific penalties or enforcement measures in the statement.

This initiative emerges amid intensifying AI competition between the United States and China. U.S. officials have emphasized advanced AI systems as crucial for national security, corporate productivity, and future economic dominance.