Anthropic has accused three Chinese AI firms of quietly extracting knowledge from its Claude model by creating thousands of fake user accounts. The company says the activity was not minor experimentation, but a coordinated effort that generated millions of interactions with its system.
The allegations come at a time when Washington is already debating how far it should go in restricting advanced AI chip exports to China. Anthropic argues that what it uncovered strengthens the case for tighter controls.
According to the company, the firms — DeepSeek, Moonshot AI, and MiniMax — set up more than 24,000 accounts and generated over 16 million exchanges with Claude. The technique allegedly used is known as “distillation,” a standard training method where developers use outputs from a larger model to train a smaller, more efficient one. While it is widely accepted when companies apply it to their own systems, it becomes controversial when used to learn from a competitor’s proprietary model.
Anthropic claims the activity focused on Claude’s most advanced features, including agent-style reasoning, tool usage, and coding abilities. The company says the goal appeared to be improving rival models by studying how Claude responds to complex prompts.
The scale varied across the three firms. Anthropic tracked more than 150,000 exchanges linked to DeepSeek that seemed aimed at strengthening reasoning logic and alignment, including responses to sensitive or policy-restricted topics. Moonshot AI allegedly generated over 3.4 million exchanges targeting areas such as coding, data analysis, computer-use agents, and computer vision. MiniMax accounted for roughly 13 million interactions, with activity focused heavily on coding and orchestration tasks. Anthropic says it observed MiniMax shifting significant traffic toward Claude shortly after a new model version was released.
DeepSeek drew global attention last year with the release of its open-source R1 reasoning model, which delivered competitive performance at comparatively low cost. The company is reportedly preparing to launch DeepSeek V4, which some expect could challenge leading Western systems in coding benchmarks. Moonshot AI recently introduced its open-source Kimi K2.5 model and a coding-focused agent, adding to the intensifying competition.
The dispute unfolds against the backdrop of ongoing U.S. policy debates over chip exports. Last month, the Trump administration formally permitted American companies, including Nvidia, to export certain advanced AI chips such as the H200 to China. Supporters say exports benefit U.S. businesses, while critics argue they could accelerate China’s AI progress during a critical phase of global competition.
Anthropic believes the alleged extraction effort required substantial computing resources, likely powered by advanced chips. In its public statement, the company said limiting access to high-end hardware would not only slow direct model training but also reduce the scale of large distillation campaigns. It called for closer coordination between AI developers, cloud providers, and policymakers.
Cybersecurity expert Dmitri Alperovitch, chairman of the Silverado Policy Accelerator and co-founder of CrowdStrike, said the accusations are consistent with concerns that Chinese AI progress may partly rely on learning from U.S. frontier systems. He argued that the claims strengthen the case for stricter export restrictions.
Anthropic also raised safety concerns. The company says its models are built with safeguards designed to prevent misuse in areas such as bioweapons research or malicious cyber operations. Systems trained through unauthorized distillation, it warned, may not carry over those protections, potentially increasing security risks.
As AI competition intensifies globally, the episode underscores how technical innovation, intellectual property, and geopolitics are becoming increasingly intertwined. The outcome of these debates could shape not only market leadership in AI, but also how safely and responsibly the technology evolves.
Also Read: ServiceNow Partners With Anthropic to Roll Out Claude to 29,000 Employees
