Anthropic accuses Chinese AI labs of mining Claude as US debates AI chip exports

Anthropic has brought serious accusations against three Chinese AI firms, claiming they created over 24,000 fraudulent accounts to harness its Claude AI model for their own advancements. The companies in question—DeepSeek, Moonshot AI, and MiniMax—are said to have generated upwards of 16 million interactions with Claude using a method known as "distillation." This technique is designed to replicate the strengths of established AI models, and Anthropic believes these labs specifically targeted Claude's unique abilities in areas such as agentic reasoning, tool utilization, and coding. The timing of these allegations coincides with ongoing discussions regarding the enforcement of export restrictions on advanced AI chips, aimed at limiting China's progress in this field. Distillation is frequently employed by AI developers to create more compact and cost-effective versions of their models. However, it poses a risk of enabling competitors to effectively "copy" the work of other labs. Earlier this month, OpenAI issued a statement to lawmakers, accusing DeepSeek of similar practices. DeepSeek gained attention last year for its open-source R1 reasoning model, which reportedly rivaled the performance of American frontier labs at a significantly lower cost. The company is on the verge of launching its latest model, DeepSeek V4, which is anticipated to outperform both Anthropic’s Claude and OpenAI’s ChatGPT in coding capabilities. Anthropic's analysis revealed a varied scope of interactions from these companies. For instance, DeepSeek reportedly engaged in over 150,000 dialogues aimed at refining foundational logic and alignment, particularly focusing on censorship-safe alternatives to sensitive inquiries. Moonshot AI's activity included more than 3.4 million exchanges targeting agentic reasoning, coding, data analysis, and computer vision after the introduction of its new open-source model, Kimi K2.5. MiniMax, on the other hand, concentrated 13 million interactions on coding and tool orchestration, with Anthropic claiming it observed MiniMax redirecting significant traffic to exploit Claude’s capabilities upon its release. In response to these developments, Anthropic is committed to enhancing its defenses against such distillation attacks and is calling for a unified approach from the AI industry, cloud service providers, and policymakers. The issue of American chip exports to China remains contentious, with the Trump administration recently easing restrictions on companies like Nvidia for exporting advanced AI chips, which critics argue could bolster China's capabilities significantly amid the global AI race. Anthropic emphasized that the scale of operations by DeepSeek, MiniMax, and Moonshot necessitates access to high-powered chips, reinforcing the argument for stringent export controls. Dmitri Alperovitch, chairman of the Silverado Policy Accelerator think tank and co-founder of CrowdStrike, expressed that these incidents highlight the ongoing theft of U.S. AI innovations and emphasize the need to restrict chip sales to these companies. Furthermore, Anthropic warns that distillation not only poses a threat to American AI competitiveness but also raises national security concerns. The company asserts that their systems are designed to mitigate risks associated with malicious AI applications, such as bioweapons development or cyber threats. Models created through unauthorized distillation are unlikely to uphold these safeguards, potentially leading to the proliferation of dangerous capabilities without adequate protections. The risk is exacerbated when these models are made open-source, making them accessible to authoritarian regimes for purposes like cyber operations and mass surveillance. TechCrunch has reached out to DeepSeek, MiniMax, and Moonshot for their responses regarding these allegations.