AI Model Theft: Anthropic Accuses Chinese Labs of ‘Distillation Attacks’

Anthropic has publicly accused three China-based AI companies – DeepSeek, Moonshot AI, and MiniMax – of orchestrating large-scale campaigns to extract capabilities from its Claude AI model through the use of tens of thousands of fraudulent accounts. The company alleges these actions violate its terms of service and raise concerns about intellectual property and competitive advantage in the rapidly evolving AI landscape.

What are ‘Distillation Attacks’?

The technique at the center of the accusations is known as “distillation.” While distillation is a common practice in AI development – where a “teacher” model is used to train a “student” model – Anthropic distinguishes these actions as “distillation attacks.” According to Anthropic, these attacks involve prompting AI models with tailored inputs and using the responses to improve other models, essentially copying the work of the original developer. TechCrunch reports that this is a method competitors can use to essentially copy the homework of other labs.

The Scale of the Accusations

Anthropic claims the three labs collectively generated over 16 million exchanges with Claude through approximately 24,000 fake accounts. The scale of the alleged activity varied between companies:

• MiniMax: Over 13 million exchanges, representing the largest effort.
• Moonshot AI: Over 3.4 million exchanges, focusing on agentic reasoning, tool use, coding, and data analysis. TechCrunch notes that Moonshot recently released a fresh open source model, Kimi K2.5, and a coding agent.
• DeepSeek: Approximately 150,000 exchanges, aimed at improving foundational logic, and alignment.

Broader Implications and US Export Controls

These accusations arrive as U.S. Officials debate the implementation of export controls on advanced AI chips, a policy intended to slow China’s AI development. Anthropic argues that circumventing regional use restrictions and breaking rules allows “foreign labs, including those subject to the control of the Chinese Communist Party, to close the competitive advantage that export controls are designed to preserve through other means.” TechCrunch highlights this connection.

OpenAI’s Similar Concerns

Anthropic is not alone in raising concerns about this practice. Earlier in February, OpenAI accused DeepSeek of similar “free-riding” on the capabilities developed by U.S. AI labs. VentureBeat reported on this memo sent to House lawmakers.

DeepSeek’s Rising Profile

DeepSeek gained attention last year with the release of its open-source R1 reasoning model, which reportedly matched the performance of American frontier labs at a fraction of the cost. The company is expected to release DeepSeek V4 soon, with reports suggesting it could outperform both Claude and ChatGPT in coding. TechCrunch warns that this release could cause volatility in the market.

Key Takeaways

• Anthropic accuses DeepSeek, Moonshot AI, and MiniMax of using over 24,000 fake accounts to “distill” capabilities from its Claude model.
• The alleged activity involved over 16 million exchanges with Claude.
• The accusations highlight concerns about AI model theft and the effectiveness of U.S. Export controls.
• OpenAI has also accused DeepSeek of similar practices.
• DeepSeek is poised to release a new model, V4, which could challenge leading AI systems.