Anthropic Accuses Chinese Labs of Claude Data Mining as US Debates AI Chip Exports

Leading AI firm Anthropic has accused three Chinese artificial intelligence companies of establishing over 24,000 fake accounts via its Claude AI model to enhance their own proprietary systems.

The implicated laboratories—DeepSeek, Moonshot AI, and MiniMax—allegedly generated more than 16 million interactions using a technique known as "distillation." According to Anthropic, these entities specifically targeted Claude's most distinct capabilities, including agent reasoning, tool utilization, and coding.

These allegations emerge amidst ongoing debates regarding the strict enforcement of export controls on advanced AI chips, a policy designed to curb China's artificial intelligence development.

Distillation is a common training methodology used by AI labs to create smaller, cost-effective model versions; however, competitors can exploit it to essentially replicate another laboratory's achievements. Earlier this month, OpenAI sent a memorandum to House representatives accusing DeepSeek of utilizing distillation to mimic their products.

DeepSeek first gained significant attention a year ago with the release of its open-source R1 reasoning model, which matched the performance of leading American labs at a fraction of the cost. DeepSeek is expected to soon launch its latest model, DeepSeek V4, which reportedly surpasses Anthropic's Claude and OpenAI's ChatGPT in coding capabilities.

The scale of each attack varied. Anthropic traced over 150,000 interactions originating from DeepSeek, which appeared aimed at improving foundational logic and alignment, particularly regarding safe alternatives for censorship around policy-sensitive queries.

Moonshot AI accounted for more than 3.4 million interactions, targeting agent reasoning, tool use, coding, data analysis, computer-use agent development, and computer vision. Last month, the company released a new open-source model, Kimi K2.5, along with a coding agent.

MiniMax conducted 13 million interactions focused on agent coding, tool usage, and orchestration. Anthropic stated it could observe MiniMax's actions because nearly half of its traffic was redirected to extract capabilities upon the release of Claude's latest model.

Anthropic indicated it will continue investing in defenses to make distillation attacks harder to execute and easier to identify, while calling for a "coordinated response among the AI industry, cloud service providers, and policymakers."

These distillation attacks occurred while the export of US chips to China remains heavily debated. Last month, the Trump administration formally allowed US companies like Nvidia to export advanced AI chips, such as the H200, to China. Critics argue that loosening these export controls boosts China's AI computing power at a critical moment in the competition for global AI dominance.

Anthropic noted that the scale of extraction performed by DeepSeek, MiniMax, and Moonshot "requires access to advanced chips."

"Thus, distillation attacks reinforce the rationale for export controls: limiting chip access not only restricts direct model training but also caps the scale of illegal distillation," according to Anthropic's blog.

Dmitri Alperovitch, Chairman of the Silverado Policy Accelerator think tank and co-founder of CrowdStrike, told TechCrunch that he was not surprised by these attacks.

"It is clear that part of the rapid progress of Chinese AI models is due to stealing US frontier models through distillation. Now we know this is a fact," Alperovitch said. "This should give us stronger reasons to refuse to sell any AI chips to these [companies], as selling only further benefits them."

Anthropic also stated that distillation not only threatens to undermine US AI dominance but may also pose national security risks.

"Systems built by Anthropic and other US companies prevent state and non-state actors from using AI to, for example, develop biological weapons or conduct malicious cyber activities," wrote Anthropic's blog post. "Models built through illegal distillation are less likely to retain these safeguards, meaning dangerous capabilities could proliferate with many protections completely stripped away."

Anthropic pointed out that authoritarian governments deploy frontier AI for things like "offensive cyber operations, disinformation campaigns, and mass surveillance," and if these models are open-sourced, the risk multiplies.