Anthropic Accuses Alibaba of Using 25,000 Fake Accounts to Extract AI Model Capabilities

Taipei, June 25 (CNA) U.S. artificial intelligence (AI) startup Anthropic recently sent a letter to the U.S. Congress accusing China's Alibaba of illegally extracting the core capabilities of its AI model, Claude, through nearly 25,000 fake accounts. This is described as the largest-scale "distillation attack" the company has ever encountered.

According to a report by Sing Tao Daily, as disclosed in Anthropic's letter to the U.S. Senate Banking Committee on June 10, operators associated with Alibaba and its AI lab "Qwen" used nearly 25,000 fake accounts between April 22 and June 5 to interact with its Claude model approximately 28.8 million times. This was done to circumvent usage restrictions and collect the output content of the Claude model.

In its letter, Anthropic stated that this attack targeted Claude's advanced capabilities in areas such as software engineering, agentic reasoning, and long-range tasks, which are the company's most commercially valuable functions.

The report mentions that a "distillation attack" involves using the output results of a more powerful and expensive AI model to train a smaller model, thereby reducing development costs. This technique allows developers to replicate some of the capabilities of advanced AI models at a lower cost. Alibaba allegedly used this unauthorized method to obtain the results of U.S. AI models, repackage them as their own products, and shorten research and development time while lowering training costs.

Prior to Alibaba being accused by Anthropic, Chinese AI startup DeepSeek had also been repeatedly accused of extracting AI models from U.S. companies in this manner.

In its letter, Anthropic warned that if Chinese AI companies bypass restrictions through a large number of accounts to obtain cutting-edge U.S. AI model capabilities, it will accelerate China's progress in AI technology. The company urged the Donald Trump administration to implement stricter technical and regulatory measures to prevent U.S. AI technology from being massively extracted by competitors.

The report notes that the U.S. Congress is considering legislation on this matter. Republican Senator Bill Hagerty and Democratic Senator Andy Kim plan to propose an amendment authorizing the U.S. government to blacklist or impose sanctions on Chinese companies that use the output content of U.S. AI models to train competing products.

Notably, the U.S. Department of Defense added Alibaba to its "Chinese Military Companies" blacklist on June 8. Alibaba has since filed a lawsuit in a U.S. court seeking to overturn this designation. (Editor: Chiu Kuo-chiang / Tien Rui-hua) 0625

Stand with facts, your every sponsorship is a force to protect the freedom of the press

Download the CNA "One-Stop News" APP for real-time updates

The text, images, and videos on this website may not be reproduced, publicly broadcast, publicly transmitted, or used without authorization.