Anthropic’s accusation against Alibaba matters because it turns an abstract AI term into a concrete business conflict. In a June 10, 2026 letter later reported by Reuters and Bloomberg, Anthropic alleged that operators linked to Alibaba and its Qwen lab used nearly 25,000 fraudulent accounts to generate about 28.8 million exchanges with Claude between April 22 and June 5, 2026. Anthropic called it the largest distillation attack it has detected so far. The company also says it does not offer commercial access to Claude in China, which means the alleged operation depended on fake identities and attempts to bypass regional restrictions. As of June 26, 2026, the reporting reviewed here did not show a public Alibaba response to the specific claim. (investing.com)

The key point is that “distillation” itself is not automatically illegal or unethical. Anthropic says frontier labs often distill their own large models into smaller, cheaper ones, and Google has likewise described distillation as a legitimate machine-learning technique when done with permission. The controversy begins when a rival allegedly harvests outputs from another company’s model at industrial scale, using those answers to train a competing system without paying the original research and computing costs. Anthropic has framed this as a direct threat to both commercial advantage and model safety, because copied capabilities may travel without the original safeguards. (anthropic.com)

So is this a new form of industrial espionage? In one sense, yes. On April 23, 2026, the White House said foreign entities, mainly based in China, were conducting “deliberate, industrial-scale campaigns” to distill U.S. frontier AI systems, and promised more coordination and possible accountability measures. A recent CNAS analysis goes further, defining “adversarial distillation” as unauthorized extraction of model capabilities through tactics such as fraudulent credentials and geographic evasion. In other words, nobody is stealing a safe full of blueprints; instead, the alleged theft happens through millions of prompts and replies. The method is new, but the motive is familiar: gain a competitor’s know-how faster and more cheaply than building it yourself. That is why this story feels less like ordinary scraping and more like the digital-age successor to industrial spying. (whitehouse.gov)