Anthropic has accused Alibaba of illicitly extracting capabilities from its Claude models in what the AI lab called the largest known attack of its kind against the company, according to a letter seen by Reuters. The campaign ran from April 22 to June 5, 2026, and generated more than 28.8 million exchanges with Claude through nearly 25,000 fraudulent accounts -- a scale that dwarfs prior incidents.
The alleged technique is distillation: systematically querying a stronger model and using its outputs to train a less capable one, effectively cloning hard-won capabilities at a fraction of the cost. Anthropic said in the letter that the effort was designed to help accelerate China's path toward the advanced 'Mythos Preview' capabilities Anthropic is developing -- framing it not as ordinary scraping but as a targeted attempt to short-circuit years of frontier research.
The accusation lands on a fault line that has been widening all year. In February, Anthropic said it had identified a campaign by Chinese AI startup DeepSeek and two other Chinese labs to extract Claude capabilities, though those operations involved far fewer exchanges. The Alibaba claim suggests both that the practice is intensifying and that the biggest Chinese tech players -- not just upstarts -- may be involved, even as Alibaba's Qwen models earn genuine praise for original research.
“The accusation lands on a fault line that has been widening all year.”
The context is a US-China AI rivalry that has so far centered on hardware: export controls on Nvidia GPUs, the fight over ASML's lithography machines, and Washington's effort to choke China's access to advanced chips. Anthropic's allegation moves the battlefield to the models themselves, raising the prospect that capability theft -- not just compute denial -- becomes the next front. It also complicates the open-weight debate, since distillation is far easier against API-accessible frontier models than against the open models Chinese labs increasingly ship.
For founders and investors, the episode is a warning about how fragile the frontier moat may be. If a leading lab's best capabilities can be partially reconstructed through 29 million well-designed queries, the durability of model-based advantages -- and the valuations built on them -- comes into question. It also strengthens the case for the labs to lock down API access, watermark outputs and pursue legal and policy remedies.
The bear case for Anthropic's framing: distillation is notoriously hard to prove, the line between legitimate benchmarking and illicit extraction is blurry, and a letter is not a lawsuit. Alibaba has strong incentives to dispute the characterization, and independent verification of the claim is not yet public. What to watch: whether Anthropic escalates to legal action or pushes for policy intervention, how Alibaba responds, and whether other frontier labs disclose similar campaigns -- the tell that capability theft has become a systemic risk to the whole industry.