Anthropic claims that Alibaba is conducting the largest distillation campaign targeting Claude.

      TL;DR: Anthropic has accused Alibaba’s Qwen lab of utilizing 25,000 fake accounts to conduct nearly 29 million exchanges with its AI model, Claude, making it the largest distillation campaign against a US AI firm yet. The accusation was communicated to senators and White House officials, highlighting the significant scale of the campaign that occurred from April to June, focusing on software engineering and agentic reasoning, which are vital commercial skills of Claude. This is the first time Anthropic has pointed to a major Chinese tech giant as the perpetrator of a distillation attack, previously citing smaller Chinese AI startups. Distillation involves using specially formulated queries to extract information from an AI model, which is then used to train a cheaper imitation. The White House has designated this method a national security threat, and after a memo from OSTP Director Michael Kratsios aimed to increase awareness of such actions, Anthropic noted that the Alibaba campaign continued regardless.

      Alibaba declined to comment on the accusations, while Anthropic stressed the need for collaboration between government and industry to combat distillation. Following the news, Alibaba's American depositary receipts fell over three percent, a decline that compounds the challenges the company already faces in the US. The Pentagon recently added Alibaba to a blacklist of Chinese military companies, to which Alibaba has responded by suing the Defense Department for removal, arguing that the designation is unfounded. The distillation allegations frame Alibaba not only as a business with potential military connections but also as actively participating in what Anthropic describes as the systematic theft of American AI technologies.

      In its letter, Anthropic cautioned that adversarial distillation enables Chinese labs to replicate advanced AI models at a reduced training cost, often lacking essential safety features. The firm urged the administration to clarify antitrust regulations to facilitate greater information sharing among US labs regarding distillation attempts and called for penalties against companies employing such tactics. Senators Bill Hagerty and Andy Kim are set to propose an amendment to defense legislation targeting any Chinese firms found improperly accessing US AI model outputs, while a related bipartisan bill in the House is under consideration, though its fate in the final defense bill remains uncertain.

      The situation is particularly timely for Anthropic, valued at $965 billion following a recent funding round, as it has confidentially filed for an IPO and is preparing for a potential listing this autumn. US officials have estimated that unauthorized distillation incurs billions in losses for Silicon Valley, posing a significant risk as cheaper Chinese alternatives threaten to undermine customer bases. However, Anthropic's calls for government intervention may not be entirely welcomed, as the company is currently engaged in a separate dispute with the Trump administration regarding export controls on its Fable 5 and Mythos 5 models, which were recently restricted due to security concerns.

      This predicament puts Anthropic in a challenging position, needing government action against Chinese competitors while also contesting restrictions placed on its own products. The letter to lawmakers seeks to delineate these issues, positing that defending US models from distillation and allowing their commercial use are not mutually exclusive aims. The response from Washington will significantly influence the regulatory landscape for US AI enterprises as well as the competitive dynamics within the industry, particularly as Anthropic has now identified four Chinese labs, with Alibaba being the largest offender in terms of scale. If legislative measures move forward, their implications could extend beyond just Anthropic to the broader question of how the US enforces protections around AI intellectual property, which can be easily replicated over the internet through sophisticated prompts.