Anthropic has accused Alibaba’s Qwen lab of running what it calls the largest distillation attack in its history: a coordinated campaign that allegedly used nearly 25,000 fake accounts to generate 28.8 million conversations with Claude, all in an effort to copy the model’s capabilities and train a rival AI. The claims were laid out in a letter Anthropic sent to the US Senate Banking Committee on June 10, 2026, which became public around June 24, and they have turned an obscure technical practice into a national-security-flavored fight over intellectual property.
The technique at the center of the accusation is called distillation. In plain terms, you point a cheaper model at a stronger one, ask it millions of questions, and harvest the answers at scale to train your own system to imitate the leader. According to Anthropic, the campaign ran from roughly April 22 to June 5, 2026, and specifically targeted Claude’s most valuable capabilities: advanced software-engineering skills and multi-step agentic reasoning, the long-horizon planning that makes a model useful for complex, autonomous tasks. Those are precisely the abilities that are hardest to build and most commercially valuable to copy.
The scale is what makes the story land. Tens of thousands of fraudulent accounts and nearly 29 million queries is not a hobbyist scraping data; it is an industrial operation. Anthropic frames it as evidence that defending a frontier model has less to do with patents and more to do with fraud detection at scale. You cannot patent the behavior of a neural network in any practical way, so the competitive moat is operational: catching and blocking the coordinated abuse of your own API before a competitor can siphon out the very capabilities you spent hundreds of millions of dollars to create.
There is a sharp irony that even Anthropic’s framing leans into. A 28.8 million query operation aimed at a single model is, as one analysis put it, a backhanded valuation of that model. You do not mount an industrial-scale copying campaign against a system you consider mediocre. In that sense, the alleged attack is a kind of admission about where the copier believes it actually stands in the race. Distillation only ever captures a snapshot, and the leading model keeps advancing while the copy chases a frozen image of last quarter’s capabilities. A copy is a lagging indicator of leadership, not a way to seize it.
For sales teams tired of cold leads, slow customer responses, and manual processes, Dapta is the ultimate tool.
Dapta is the leading platform for creating AI sales agents specifically designed to increase inbound lead conversion. Respond to your leads in less than a minute with voice AI and WhatsApp that converts.
If you want your team to sell more while AI handles the complex stuff, you have to try it.
For companies outside the AI labs, the episode is a concrete lesson in a risk that is easy to ignore: your proprietary data and outputs can leak through the very interfaces you expose to the world. If your business generates valuable content, structured knowledge, or model outputs, and you expose them through an API or a product, you are potentially handing competitors the raw material to imitate what makes you distinctive. The competitive edge embedded in a system can drain out one query at a time, and the defense is the unglamorous work of monitoring usage, detecting coordinated abuse, and enforcing terms at scale.
The geopolitical subtext is impossible to miss. The accusation fits a broader narrative about Chinese labs closing the gap with US frontier models, sometimes through genuine research and sometimes, as Anthropic alleges here, through copying the leader’s answers. Whether or not the specific claims hold up, and Alibaba has not addressed the details, the strategic point stands: in a world where training a competitive model from scratch costs a fortune, copying the output of the leader is a tempting shortcut. China does not need to win the training race if it can buy the result for 28 million conversations. That is the uncomfortable reality this letter forced into public view, and it is why AI intellectual property is quickly becoming a matter for Congress, not just corporate legal teams.
