Anthropic, an artificial intelligence company, has formally accused Chinese technology and e-commerce leader Alibaba of illicitly extracting capabilities from its Claude AI model. According to a letter seen by Reuters, Anthropic described this event as the largest known distillation attack on the company to date. Distillation is a technique where a smaller, weaker model is trained on the outputs generated by a larger, more powerful model to replicate its capabilities.
This accusation is part of a broader pattern of alleged intellectual property theft from Anthropic. The company also claims that three other Chinese AI firms — DeepSeek, Moonshot AI, and MiniMax — used 24,000 fraudulent accounts to siphon off Claude's capabilities, generating more than 16 million exchanges. These labs reportedly targeted advanced reasoning, coding, and tool-use capabilities, with DeepSeek specifically focusing on generating 'chain-of-thought' training data over 150,000 exchanges. Moonshot AI engaged in 3.4 million exchanges for reasoning and coding, and MiniMax was the heaviest user with 13 million exchanges.
All three labs allegedly accessed Claude through proxy services, circumventing Anthropic's restrictions in China where it does not operate commercially. This comes after similar accusations from OpenAI and Google, which reported distillation attacks on its Gemini model involving over 100,000 prompts. The ability of Chinese labs to replicate Claude's features through such methods could erode Anthropic's competitive advantage and impact its substantial valuation, which recently reached $380 billion after a $30 billion Series G funding round. Major investors like Amazon, which has invested $8 billion, and Palantir, which integrates Claude into classified defense environments, could be affected by these developments.
