The June 10 letter to US senators alleges Alibaba's AI lab, Qwen, automated 28.8 million queries through nearly 25,000 fake accounts to harvest Claude's answers as training data.
A US AI lab has publicly accused a Chinese AI lab of model theft. That framing suggests a heist: weights copied, server breached, secrecy broken. Anthropic's June 10 letter to US senators alleges something different in kind, not just degree (CNBC). What Anthropic describes is not the theft of Claude itself, but the systematic harvesting of its answers through fake accounts: about 25,000 accounts, roughly 28.8 million exchanges with Claude between April 22 and June 5, 2026.
The technique is called distillation. In plain terms, it is copying the behavior of a powerful AI by querying it heavily and using its answers to teach a smaller or rival system. Distillation is the dominant way AI labs learn from each other in 2026, and a normal amount of it, conducted through licensed APIs and traced to identifiable customers, is part of the industry's plumbing. Anthropic's allegation is that this batch fell outside that normal range. The accounts are disposable. The prompts are scripted. The exchange rate is dense enough to look like an industrial operation rather than research curiosity.
The mechanism matters because it does not require stealing Anthropic's model weights, the actual mathematical recipe that makes Claude respond the way it does. It works entirely through Anthropic's own chat interface and API, the public endpoints any paying customer can call. Whoever ran the campaign stayed inside the gate Anthropic had built, then walked out with a thick stack of labeled examples. That distinction is what most public commentary blurs. Headlines that say 'AI theft' suggest an equivalent to the theft of the model itself, while Anthropic's own language, describing Alibaba's behavior as 'brazenly' illegal and the largest such case it has publicly disclosed, leans into that equivalence.
The 28.8 million exchange figure and the 25,000-account count are sourced to Anthropic's letter. The BBC and TechTimes confirmed the letter's existence and reported the same specifics (BBC, TechTimes); secondary aggregation tracks Bloomberg as the first outlet to report (MemeBurn). No bundled source provides independent technical verification of those numbers. The characterizations 'brazenly' and 'largest case' are Anthropic's framing, not audited fact, and downstream reporting should treat them that way.
Anthropic's parallel technical post describes the detection methods and rate limits it uses to identify this kind of abuse (Anthropic: detecting and preventing distillation attacks). That post is the practical answer to 'what does Anthropic actually do about it.' The letter itself is the political answer: the company is asking for federal rules on API misuse, on the grounds that no single AI lab can police this behavior on its own. The argument is reasonable on its face. It is also the argument of a company with a direct commercial and policy interest in framing API-level learning as theft.
Market reaction so far: Alibaba's stock is sharply down year-to-date in 2026 and reportedly hit a 16-month low in the days after the letter. The exact trading day and exchange, NYSE for the ADR or HKEX for the primary listing, should be pinned down before any version of the 16-month figure is repeated as fact.
What is missing. No bundled source contains an Alibaba on-record response: no denial, no technical rebuttal, no acknowledgment. In a normal corporate-IP dispute, the response carries roughly as much signal as the accusation. The documented absence of an Alibaba statement is itself part of this story, and any substantive Alibaba response on Qwen training would reset the framing on contact.
What to watch next. A first on-record Alibaba statement on Qwen's training pipeline. An SEC or HKEX filing on the same subject, which would shift the dispute from rhetoric to enforcement. Independent confirmation or refutation of Anthropic's exchange count, whether from a third-party researcher or a court filing. And a federal bill naming API misuse as a regulatory category, which would mark the political ask landing.