Anthropic, OpenAI, and Google are sharing threat intelligence after alleging Chinese labs used millions of fake interactions to clone their AI models on the cheap
Think of it like photocopying someone else’s homework, except the homework cost billions of dollars to produce and the photocopier is running at industrial scale. That’s essentially what major US artificial intelligence companies are alleging their Chinese competitors have been doing, using a technique called “distillation” to systematically extract the intelligence baked into America’s most advanced chatbots.
Anthropic, the maker of the Claude AI model, disclosed in a February 23, 2026, blog post that three Chinese labs, DeepSeek, Moonshot AI, and MiniMax, had collectively created roughly 24,000 fake accounts. Those accounts were used to conduct over 16 million conversations with Claude, harvesting its responses to train competing systems at a fraction of the original development cost.
How distillation works and why it matters
Distillation, in its legitimate form, is a well-known machine learning technique. A smaller, cheaper model learns to mimic the outputs of a larger, more expensive one. It’s how companies compress powerful AI into something that can run on a smartphone.
The problem arises when the “teacher” model doesn’t belong to you. These Chinese labs allegedly treated Claude and OpenAI’s models like an all-you-can-eat buffet, sending millions of queries through fake accounts, recording the responses, and feeding that data into their own systems.
OpenAI has leveled its own parallel accusations. The company told the House Select Committee on China in early February 2026 that DeepSeek had employed distillation tactics to develop its R1 model.
The allegations escalated further in June 2026, when new accusations emerged targeting Alibaba. That attack has been described as the largest known distillation operation to date, involving approximately 28.8 million interactions with AI models.
The US response: intelligence sharing and White House involvement
The scale of these alleged operations prompted something unusual in Silicon Valley: cooperation. OpenAI, Anthropic, and Google began sharing threat intelligence starting in April 2026, pooling information about attack patterns, fake account networks, and the methods being used to circumvent their systems’ safeguards.
A White House memo dated April 23, 2026, confirmed that the administration views these distillation campaigns as “industrial-scale” operations by Chinese companies.
