Anthropic accuses Alibaba of the largest known distillation attack on Claude

Anthropic has levelled its most serious anti-distillation accusation to date at a Chinese rival — and it did so in front of Congress. In a 10 June 2026 letter to the US Senate Banking Committee, first reported by Bloomberg, the company alleged that operators affiliated with Alibaba and its AI lab ran what it calls the largest known model-distillation campaign against Claude: roughly 28.8 million model exchanges through some 25,000 fraudulent accounts over a 44-day window from 22 April to 5 June 2026. Two things should anchor the reader before any analysis: these are Anthropic's allegations, and Alibaba denies them. None of the figures has been independently verified.

By Anthropic's account, the campaign was not opportunistic scraping but a targeted extraction, concentrated on Claude's software-engineering, agentic-reasoning and cybersecurity capabilities — and specifically on the frontier model Anthropic refers to as Mythos Preview. Those are the highest-value, hardest-to-replicate skills a lab has, which is why the choice of target matters as much as the volume. Senator Elizabeth Warren, citing the letter, described the episode as 'the largest known distillation attack on Anthropic to date'. The venue is itself a signal: Anthropic routed the claim through a Senate committee rather than a blog post, folding it into a week already thick with export-control pressure on the company.

Model distillation means training a smaller, cheaper model on the outputs of a more powerful one so the student imitates the teacher. Done with permission it is ordinary engineering; done covertly against a competitor's paid API it is, in Anthropic's telling, theft of the most expensive thing a frontier lab owns. Anthropic's sharper argument is not commercial but about safety: a model distilled from Claude can pick up the raw capability while leaving behind everything Anthropic layered on top of it — the safety fine-tuning, the usage policies, the refusals, the access controls. The skill is copied; the brakes are not.

This is not Anthropic's first such accusation. In February 2026 the company said it had identified three separate 'industrial-scale' distillation campaigns, which it attributed to DeepSeek, Moonshot and MiniMax. What makes the Alibaba allegation land harder is scale: Anthropic says the ~28.8-million-exchange campaign is larger than that entire earlier set combined. Read charitably, that is a company documenting an intensifying pattern of capability theft by Chinese labs. Read sceptically, it is a well-timed escalation — the disclosure arrived the same week Anthropic was under acute US export-control pressure, and a vivid 'largest ever' number is useful to a company arguing that Washington should treat its models as strategic assets.

The honest position, as of now, is that both things can be true and neither is proven. The 28.8-million exchange count and the ~25,000 account figure are Anthropic's allegations; Alibaba rejects them; no independent party has confirmed either number. What is verifiable is that a leading US lab has publicly accused a leading Chinese one of large-scale capability theft, in a formal letter to Congress, during a fight over how tightly the US should control frontier models. That combination — serious accusation, disputed evidence, high political stakes — is the story, and it will not resolve on the strength of a headline number — the burden of proof sits with Anthropic, and so far it has offered assertion rather than evidence.