Anthropic says Alibaba used tens of thousands of fake accounts to pull information from its Claude chatbot in order to improve the Chinese company’s own AI models. In a letter sent to White House officials and senators, Anthropic said the effort was carried out by Alibaba’s AI lab, Qwen, and focused on Claude’s most valuable abilities, including software engineering and AI agents.
The company called it the largest attempt it has identified by a Chinese firm to improperly exploit an American AI model. Anthropic said the pattern fits a broader trend of Chinese companies using unauthorized responses from U.S. models to develop rival systems at a fraction of the cost, while ending up with models that lack the safety protections found in American systems.
Anthropic cited an earlier case in February, when it said DeepSeek and two other Chinese firms had linked to Claude through fake accounts to carry out “distillation,” a process in which a less advanced model is trained on the outputs of a stronger one. In that case, the three firms made 16 million interactions with Claude using 24,000 fake accounts. Alibaba’s activity was larger, Anthropic said, involving 25,000 accounts and 28.8 million interactions between April and June.
The company warned, “Distillation attacks are carried out illegally, systematically and at industrial scale in order to harvest American AI capabilities from frontier labs, and repackage them as their own without the training and R&D costs involved in developing American frontier models.” Anthropic added, “Distillation must be fought with coordinated government and industry action.”
The letter is already prompting action in Washington. Senators Bill Hagerty, a Republican, and Andy Kim, a Democrat, have introduced a bill amendment that would blacklist or sanction any Chinese company found to improperly use an American AI model to train its own systems. A similar amendment has been introduced in the House, but it is not yet clear whether either measure has enough support to pass.
