Anthropic accused DeepSeek, Moonshot AI, and MiniMax of creating over 24,000 fake accounts to extract capabilities from Claude. The companies allegedly generated over 16 million exchanges focused on reasoning, coding, tool use, and computer vision. Anthropic described the activity as distillation, a training method that becomes competitive copying when one lab uses another model's answers as a source. The accusations arrived while US officials debated controls on advanced AI chips exported to China, TechCrunch reported on February 23rd.

The model is presented as a sealed product, but its behavior can be sampled one answer at a time. Thousands of accounts turn the chatbot into an unwitting tutor whose lessons are collected elsewhere and compressed into a rival system. The theft is difficult to see because the exchange resembles ordinary use until its scale reveals an industrial operation. Export controls attempt to restrict the machines that run the models. Distillation attacks target the behavior that those machines produce. A frontier model therefore becomes both a national asset and a public surface exposed through its own interface. The server room is guarded, but the most valuable material leaves as a series of answers.