Anthropic joins OpenAI in flagging 'industrial-scale' distillation campaigns by Chinese AI firms

Anthropic has recently raised alarms about three Chinese AI companies—DeepSeek, Moonshot AI, and MiniMax—claiming they have orchestrated efforts to unlawfully extract knowledge from Anthropic's model, Claude. This accusation follows similar complaints made by OpenAI, marking a growing trend among American tech firms voicing their concerns regarding international AI practices. In a statement released by Anthropic, these companies are said to have conducted significant 'distillation attack' operations. This involves bombarding Claude with a multitude of crafted prompts to train their proprietary models. The technique of distillation allows smaller AI systems to replicate the performance of larger, well-established models by siphoning off knowledge, which is particularly advantageous for teams with limited resources. Despite Anthropic implementing restrictions that prohibit commercial access to Claude in China, these firms allegedly circumvented these barriers by utilizing commercial proxy services. This approach allowed them to gain access to networks that operated tens of thousands of Claude accounts simultaneously. According to Anthropic, once access was gained, the labs produced vast amounts of tailored prompts aimed at extracting specific capabilities from Claude. The responses generated by Claude were reportedly harvested in bulk, either for direct training of the Chinese models or to facilitate a process known as reinforcement learning. This method involves AI models improving their decision-making through trial and error without human assistance. Anthropic estimates that these three companies collectively managed to generate over 16 million interactions with Claude, using approximately 24,000 accounts that were created fraudulently. Among them, MiniMax was identified as the most active, responsible for over 13 million exchanges. As of now, DeepSeek, Moonshot AI, and MiniMax have not issued any comments regarding these allegations. Anthropic's claims add to a growing number of American firms raising the alarm over the distillation practices of Chinese AI entities. Earlier in the month, Sam Altman's OpenAI submitted a letter to U.S. lawmakers, highlighting what they described as ongoing attempts by DeepSeek to distill their models using sophisticated methods. While distillation has become a common practice in the AI industry—acknowledged by Anthropic as a routine process—there are apprehensions about the competitive edge that could be gained by rival firms through such methods. Anthropic expressed concern that this practice could allow competitors to acquire significant capabilities rapidly and at a fraction of the cost it would take to develop them independently. Both Anthropic and OpenAI have framed these activities by Chinese firms as potential national security threats. They emphasized the risk of authoritarian governments leveraging advanced AI technologies for offensive operations, disinformation efforts, and mass surveillance. However, the extent to which these statements reflect genuine security concerns or a desire to maintain competitive superiority remains uncertain. Critics have also pointed out the irony in Anthropic's concerns, given their own use of distillation to develop proprietary models. Observers note that the narrative surrounding illicit capability transfer might strengthen calls for stricter export controls on advanced AI technologies to China. On the same day Anthropic made its statements, reports emerged indicating that DeepSeek had trained its model using Nvidia's Blackwell chip, allegedly violating export regulations, further intensifying scrutiny from U.S. officials. As the landscape of AI technology continues to evolve, the U.S. administration has expressed increasing anxiety over China's rapid advancements, particularly as these gains are often perceived to be achieved through American-developed systems. In a related effort, the White House recently announced the establishment of a new initiative within the Peace Corps aimed at promoting American AI interests globally and assisting partner nations in adopting advanced systems.