Anthropic is seeking help from the U.S. government in its fight against Chinese companies trying to copy its artificial intelligence models, and said Alibaba Group Holding Ltd. is among the companies behind these attacks.
The American AI company said this in a letter sent to several US senators and White House officials, Bloomberg reported on Wednesday (June 24), citing a copy of the letter.
An Anthropic spokesperson declined to discuss the reported letter with Bloomberg, but said “coordinated action between government and industry” is needed to combat distillation, the process by which developers use the results of another AI model to train their own models at a much lower cost, according to the report.
Alibaba reportedly declined Bloomberg’s request for comment. The company did not immediately respond to PYMNTS’ request for comment.
Anthropic said in the letter that Alibaba’s Kwen AI Lab used 25,000 fraudulent accounts to gain unauthorized access to its Claude AI model, conducted 28.8 million interactions with the model, and used the collected information to develop rival chatbots through adversarial distillation, Bloomberg reported.
According to the report, the company said the AI models developed through this process lack the guardrails that Anthropic has in place for its own models, and that distillation poses a threat to national security because it could allow China to narrow the US lead in AI.
Antropic also asked the senators and officials who sent the letter to clarify antitrust guidelines to allow U.S. companies to share information about distillation, support export controls for advanced AI chips, and penalize companies that use distillation, according to the report.
Anthropic said in a blog post in February that Chinese AI labs DeepSeek, MiniMax, and Moonshot AI were using Distillation and Claude’s output to train their own models.
“These campaigns are increasing in intensity and sophistication,” Anthropic said in the post. “The room for action is narrow, and the threat extends beyond a single company or region. Addressing this requires swift and coordinated action among industry players, policymakers, and the global AI community.”
In a Feb. 12 blog post, the Google Threat Intelligence Group confirmed that the incidence of distillation attacks, or “model extraction attacks,” is increasing, and said that, for example, coding models “can be targeted by attackers seeking to replicate functionality in environments without guardrails.”
