Distillation trains a weaker model on the outputs of a stronger one. It is a routine, legitimate technique; the dispute is over running it through fake accounts at industrial scale to copy a rival’s capabilities. Anthropic says the campaign focused on Claude’s software-engineering and agentic-reasoning capabilities — two of the most commercially valuable frontiers in AI — and was carried out using proxy services because Anthropic does not sell Claude commercially in China.

An escalation from February

The accusation extends a pattern Anthropic first detailed publicly on 23 February 2026, when it reported industrial-scale campaigns by three Chinese labs — DeepSeek, Moonshot and MiniMax — generating more than 16 million Claude exchanges through about 24,000 fraudulent accounts. By Anthropic’s own count, the single Alibaba operation now alleged is larger than those three combined.

Operator	Exchanges	Disclosed
DeepSeek	150,000+	Feb 2026
Moonshot AI	3.4M+	Feb 2026
MiniMax	13M+	Feb 2026
Alibaba / Qwen	28.8M	Jun 2026

In February, MiniMax ran the largest of the three campaigns at more than 13 million exchanges, focused on agentic coding; Moonshot exceeded 3.4 million, targeting agentic reasoning, tool use and computer vision; and DeepSeek topped 150,000, including prompts that turned Claude into a reward model and others that produced censorship-safe answers to politically sensitive questions. Anthropic said one proxy network alone ran more than 20,000 fraudulent accounts at once.

Why Anthropic frames it as national security

Anthropic’s argument is that illicitly distilled models inherit a rival’s capabilities without its safeguards, and that the technique lets foreign labs close the gap export controls are meant to preserve. Because running extraction at scale requires advanced chips, the company contends that distillation strengthens, rather than weakens, the case for chip export controls. It says the Alibaba operation could accelerate China’s effort to reproduce capabilities in its unreleased Mythos Preview model projection.

The letter pairs the allegation with policy asks: clearer antitrust guidance so US labs can share distillation intelligence, continued support for chip export controls, and penalties for firms that distill. Those requests are reported from the letter by a single outlet, Bloomberg, which broke the story; Reuters and CNBC later confirmed the letter’s core figures.

The Washington backdrop is uneven

Anthropic is asking for help in the same month Washington restricted its own frontier models. The US government issued an export-control directive suspending access to Anthropic’s Fable 5 and Mythos 5 for foreign nationals, leaving the company both a beneficiary and a casualty of the regime it endorses. Alibaba, meanwhile, was added this month to the Pentagon’s list of Chinese military companies, a designation it is challenging.

Enforcement elsewhere has lagged: Reuters reported this month that the Commerce Department has held off blacklisting DeepSeek despite an interagency finding that it poses a national-security risk. The picture is one of escalating accusations meeting inconsistent policy.

The case for caution

Two facts temper the allegation. First, distillation is a standard industry practice — labs distill their own models routinely — so the line between research and theft is contested. Erik Cambria, an AI professor at Nanyang Technological University, told CNBC in February that “the boundary between legitimate use and adversarial exploitation is often blurry”, and noted the disclosures may also serve to protect a US commercial lead. Second, every figure here is Anthropic’s, the June letter is not public, and Alibaba has not answered. The numbers describe what one accuser says it detected, not an audited finding.

Sources

1. Anthropic, “Detecting and preventing distillation attacks” — 23 February 2026 (per-lab figures, methodology).
2. Reuters, “Anthropic says Alibaba illicitly extracted Claude AI model capabilities” — 24 June 2026 (letter, dates, recipients).
3. CNBC, “Anthropic accuses Alibaba of campaign to brazenly and illicitly extract AI capabilities” — 24 June 2026 (28.8M / 25,000; “largest known” wording; Bloomberg first).
4. Anthropic, “Statement on the US government directive to suspend access to Fable 5 and Mythos 5” (export directive).
5. CNBC, “Anthropic, OpenAI accuse Chinese firms of distillation” — 24 February 2026 (per-lab totals; Cambria quote; OpenAI parallel).

Last updated: 25 June 2026 · First published: 25 June 2026