A Letter to Congress, Not a Lawsuit

On June 10, Anthropic sent a letter to Senators Tim Scott and Elizabeth Warren, the chair and ranking member of the Senate Banking Committee, laying out an accusation against Chinese technology giant Alibaba. The letter did not go to a courtroom. It went to Washington, ahead of a scheduled hearing on artificial intelligence. According to Reuters, which first reported the contents on June 24, Anthropic alleged that operators tied to Alibaba's Qwen AI lab ran roughly 25,000 fraudulent accounts and generated more than 28.8 million conversations with Claude between April 22 and June 5.

That choice of venue tells you something. Anthropic isn't suing a competitor. It's asking lawmakers to treat what it calls industrial-scale model extraction as a policy problem, not just a commercial dispute between two companies. Whether Congress agrees is a separate question, but the framing itself is deliberate.

What "Distillation" Actually Means

The technical term for what Anthropic alleges is distillation. A cheaper model is pointed at a stronger one, and its answers are harvested at scale, then a second model is trained to imitate those results. According to Anthropic's letter, as reported by PYMNTS, the fake accounts weren't asking Claude about the weather or making small talk. They were targeting the model's most valuable capabilities: advanced software engineering and the multi-step agentic reasoning that lets a model plan and execute long tasks on its own.

This distinction matters because distillation, by itself, is not illegal or even unusual. Companies routinely compress their own large models into smaller, faster versions for cheaper deployment. The line Anthropic is drawing is between using the technique on your own systems and running it against a rival's model without permission, at a scale designed to reverse-engineer that model's behavior. Anthropic's Head of Policy, Sarah Heck, described the alleged conduct in the letter as carried out illicitly and systematically to harvest American AI capabilities and repackage them without paying the training costs, according to Business Insider's reporting.

This Isn't Anthropic's First Accusation

Alibaba is not the first company Anthropic has named in a distillation complaint. In February 2026, Anthropic identified three other Chinese labs, DeepSeek, Moonshot AI, and MiniMax, as having collectively run about 24,000 fraudulent accounts to generate more than 16 million exchanges with Claude. Individually, the activity ranged from roughly 150,000 exchanges tied to DeepSeek up to more than 13 million tied to MiniMax, based on figures reported by Let's Data Science.

The alleged Alibaba campaign, at 28.8 million exchanges, is larger than that entire February batch combined, and it happened in roughly six weeks instead of a longer stretch. Anthropic also says the targeting sharpened between the two episodes. Where the February campaigns tested general Claude performance, the newer effort focused specifically on the coding and agentic skills that are the hardest and most expensive part of building a frontier model. If Anthropic's account holds up, the pattern isn't isolated incidents. It's an escalating and increasingly precise effort aimed at the parts of Claude that matter most commercially.

An Unverified Claim With Real Consequences Already

It's worth being direct about what this is: an accusation, not a finding. Alibaba has not publicly addressed the specifics of the claim, and no independent party has verified Anthropic's numbers. Separately, Alibaba has filed its own lawsuit seeking removal from a US government blacklist, according to Dawn's reporting on the BBC's coverage, which shows the relationship between Alibaba and Washington was already adversarial before this letter surfaced.

But the consequences moved fast regardless of verification. Two days after Anthropic's letter was sent, the US Commerce Department imposed export restrictions on Anthropic's own Mythos and Fable models, citing national security concerns, and Anthropic pulled public access to both worldwide, according to TechSpot's reporting. Those restrictions were lifted on June 30 and access to Fable 5 began restoring on July 1. The timing is difficult to read as coincidental, even though no official statement has directly linked the two events. When an AI company hands Congress a national-security-flavored accusation against a foreign rival, it should not be surprising if regulatory attention lands on that company too.

Why This Story Matters Beyond the Two Companies

For competing labs, a successful distillation campaign offers a shortcut around years of research and billions of dollars in training costs. For enterprise customers, the risk is quieter but arguably more important. Anthropic argues that a distilled model doesn't inherit the safety alignment work built into the original, meaning a copy might behave similarly on the surface while lacking the same safeguards underneath. That's not an abstract concern for anyone deploying these systems in regulated industries where model behavior under edge cases actually matters.

The broader signal here isn't really about Alibaba specifically. It's that model-theft-by-conversation is now a recognized attack category serious enough to reach a Senate committee, not just a security team's internal threat model. Anthropic has effectively told every AI lab that talking to a competitor's chatbot at scale can be treated as extraction, and that framing will likely shape how usage limits, account verification, and API terms of service evolve across the industry over the next year. Whether Alibaba did what's alleged remains unresolved. What's already resolved is that this kind of accusation now has a template, and other labs will use it again.