Anthropic has accused three Chinese AI companies of using Claude at large scale to improve their own models, turning a technical dispute over model training into a policy argument about AI chip exports.
The company says DeepSeek, Moonshot AI, and MiniMax set up more than 24,000 fake accounts and generated more than 16 million exchanges with Claude. Anthropic describes the activity as distillation aimed at Claude capabilities such as agentic reasoning, tool use, and coding.
What Anthropic Says Happened
The central allegation is that the companies used Claude as a source of training signal for their own AI systems. In Anthropic’s account, the fake accounts were not casual misuse. They were part of a coordinated effort to extract useful behavior from Claude through many exchanges.
Distillation is not unusual inside AI labs. The source describes it as a common method for using one model to create a smaller and cheaper version. The conflict begins when competitors use that method against another lab’s model, effectively trying to reproduce valuable capabilities without building them in the same way.
Anthropic says the activity focused on Claude’s most distinctive strengths. The capabilities named in the source include agentic reasoning, tool use, coding, data analysis, computer-use agent development, computer vision, and orchestration.
The company also frames the alleged activity as a security and policy issue, not only a commercial one. Its position is that large-scale extraction can help accelerate rival models while weakening the value of safeguards built into U.S. systems.
The Three Companies Named
Anthropic says the activity differed by company, both in scale and in apparent focus.
- DeepSeek: Anthropic tracked more than 150,000 exchanges that appeared aimed at improving foundational logic and alignment, including censorship-safe alternatives to policy-sensitive queries.
- Moonshot AI: Anthropic says there were more than 3.4 million exchanges targeting agentic reasoning and tool use, coding and data analysis, computer-use agent development, and computer vision.
- MiniMax: Anthropic says MiniMax generated 13 million exchanges focused on agentic coding, tool use, and orchestration.
The source also notes recent product context around the companies. DeepSeek drew major attention a year ago with its open source R1 reasoning model, which nearly matched American frontier labs in performance at a fraction of the cost. DeepSeek is expected to soon release DeepSeek V4, which reportedly can outperform Anthropic’s Claude and OpenAI’s ChatGPT in coding.
Moonshot AI released a new open source model, Kimi K2.5, and a coding agent last month. Anthropic’s MiniMax allegation includes a more specific observation: the company says it saw MiniMax redirect nearly half its traffic to pull capabilities from the latest Claude model when it was launched.
Why Distillation Matters
The dispute matters because distillation sits in a gray area between normal AI engineering and competitive extraction. When a lab uses the method on its own systems, it can make models cheaper or more efficient. When a rival uses another company’s model as the teacher, the practical effect can be much more contentious.
That is why the source compares Anthropic’s claims with OpenAI’s earlier memo to House lawmakers. OpenAI accused DeepSeek of using distillation to mimic its products. Anthropic’s claims are broader in company count and include a large volume of alleged Claude interactions.
The stakes are not limited to copying model behavior. Anthropic says U.S. companies build systems that prevent state and non-state actors from using AI for dangerous activity, including developing bioweapons or carrying out malicious cyber activities. Its concern is that models built through illicit distillation may not keep those protections.
Anthropic also points to the risk of authoritarian governments using frontier AI for offensive cyber operations, disinformation campaigns, and mass surveillance. The source says that risk is multiplied if those models are open sourced.
The Chip Export Debate
The accusations arrive as American chip exports to China remain hotly debated. Last month, the Trump administration formally allowed U.S. companies like Nvidia to export advanced AI chips, including the H200, to China.
Critics argue that loosening export controls increases China’s AI computing capacity at an important moment in the global AI race. Anthropic connects that policy debate directly to its distillation claims, saying the scale of extraction by DeepSeek, MiniMax, and Moonshot requires access to advanced chips.
Anthropic’s argument is that chip controls do not only affect direct model training. They also affect whether companies can run extraction campaigns at large scale. In the source, Anthropic says restricted chip access limits both direct model training and the scale of illicit distillation.
Dmitri Alperovitch, chairman of the Silverado Policy Accelerator think-tank and co-founder and former CTO of CrowdStrike, told TechCrunch he was not surprised by the alleged attacks. He said the claims strengthen the case against selling AI chips to the companies involved.
What Comes Next
Anthropic says it will continue investing in defenses that make distillation attacks harder to execute and easier to identify. But the company is also calling for a coordinated response across the AI industry, cloud providers, and policymakers.
That call matters because the alleged behavior crosses several boundaries at once. It touches platform abuse through fake accounts, competition between AI labs, open source model development, cloud infrastructure, and export controls.
TechCrunch has reached out to DeepSeek, MiniMax, and Moonshot for comment. For now, the source presents Anthropic’s claims as part of a broader fight over how access to frontier AI systems, advanced chips, and model safeguards should be controlled.