Anthropic says Alibaba ran a 28.8M-exchange raid on Claude. Alibaba says no.

Here is the uncomfortable part everyone rushing to retweet the number skipped: 28.8 million exchanges and 25,000 fake accounts are things Anthropic says happened, in a letter to the US Senate Banking Committee, and Alibaba says did not. No regulator has confirmed them. No independent audit has confirmed them. They are, right now, an accusation — a serious one, worth taking seriously — but an accusation dressed up in the precision of a forensic finding. Treat the decimal point with suspicion.

Notice the shape of the claim. In February, Anthropic named three Chinese labs — DeepSeek, Moonshot, MiniMax — for 'industrial-scale' distillation. Now it says the alleged Alibaba campaign is bigger than all three combined. That is a very clean escalation: each new accusation conveniently tops the last, and each arrives at a moment when Anthropic benefits from Washington seeing Chinese labs as predators and Anthropic as the victim worth protecting. The pattern might be genuine. It might also be a narrative that grows in the exact direction the company's policy interests point.

The contrarian move is not to wave the whole thing away — that would be lazy in the other direction. Strip out the disputed count and a real problem remains: distillation lets a rival train a cheap model on Claude's outputs and walk away with the capability but none of the guardrails. Claude's safety training, refusals, usage policies and access controls do not ride along in the outputs. So a distilled model can end up as capable as Claude on cyber or agentic tasks while being far freer to misuse. That is true whether the number is 28.8 million or a fraction of it.

The detail that should make any reader slow down is when this surfaced. Not in a security disclosure on Anthropic's blog, but in a 10 June 2026 letter to a Senate committee, in the same window Anthropic was fighting US export controls on its own models. A record-setting accusation against a Chinese giant is exactly the kind of exhibit that helps a US lab argue it deserves protection, not restriction. That does not make the accusation false. It does mean the accusation and the ask are entangled — and reporting that repeats the 28.8-million figure as fact, rather than as a disputed allegation, is doing Anthropic's lobbying for free.

And here is the test that would settle it, which nobody has run. A distillation claim this large is, in principle, checkable: Anthropic holds the server logs, the account-creation fingerprints, the traffic patterns and the payment trails that would show whether ~25,000 accounts really moved as one coordinated operation. If the evidence is as damning as the letter implies, the obvious move is to publish a technical write-up — anonymised but auditable, the way security teams do after a real breach. Instead we got a headline number in a note to senators. That gap between 'we have proof' and 'here is the proof' is exactly where a reader should plant their scepticism. None of this clears Alibaba — distillation of US frontier models by Chinese labs is a documented pattern. It means the specific 28.8-million figure is being asked to carry policy weight it has not yet earned, and everyone repeating it is supplying the evidence Anthropic hasn't.