Anthropic Accuses Alibaba of Mass Claude Distillation

Anthropic has accused Chinese technology giant Alibaba of running the largest campaign it has ever documented to copy its Claude models — telling U.S. senators and White House officials that operators tied to Alibaba's Qwen AI lab used roughly 25,000 fraudulent accounts to generate nearly 28.8 million exchanges with Claude between April 22 and June 5, 2026. The queries, Anthropic alleges, were engineered to harvest Claude's most valuable skills: software engineering and agentic reasoning.

The technique at issue is distillation — feeding carefully constructed prompts to a frontier model, collecting its responses, and using that data to train a cheaper rival that approximates the original's behavior. Done at scale, it lets a competitor short-circuit the hundreds of millions of dollars and years of research that went into the source model. By focusing on coding and agent tasks, Anthropic says, the operation went straight for the capabilities that are hardest to build and most lucrative to sell.

What makes this filing notable is the scale and the target. The single Alibaba campaign, by Anthropic's count, exceeded the combined volume of the three Chinese startups it named in February — DeepSeek, MiniMax and Moonshot AI — which together ran some 16 million exchanges through about 24,000 fake accounts. It is also the first time Anthropic has publicly named a major Chinese conglomerate rather than a startup. The company stressed that the activity continued after an April White House memo had explicitly flagged distillation as a national-security concern — in its framing, in open defiance of Washington's warnings.

Anthropic is using the episode to press for policy action: clearer antitrust guidance so AI firms can share information about distillation campaigns, renewed support for export controls on advanced AI chips, and penalties for companies caught using the technique. The asks land amid an already tense stretch of U.S.–China AI policy, from chip restrictions to the recent export-control directive that pulled Anthropic's own Fable 5 offline. Alibaba, for its part, declined to comment on the allegations.

The accusation is, for now, exactly that — an allegation, with no independent adjudication and a flat non-response from the accused. But it crystallizes the question increasingly at the center of the AI race: when a model's outputs can be used to clone its hardest-won abilities, how does any lab protect what it spent billions to build — and can terms-of-service violations even be policed across borders?