Anthropic has accused operators linked to Chinese tech giant Alibaba of using fake accounts to siphon knowledge out of its Claude AI models, according to multiple reports.
According to InfoWorld, Computerworld, and Forbes, Anthropic alleges the campaign relied on roughly 25,000 fraudulent accounts to access Claude. Computerworld reports the effort generated 28.8 million queries in what it described as a large-scale model-extraction attempt to replicate Claude's capabilities. WION characterized the activity as some 29 million "adversarial distillation" attacks.
The accusation centers on a practice known as distillation. In simple terms, this means repeatedly querying a powerful AI model and using its answers as training material to build a cheaper, competing model — effectively letting one company learn from the expensive work of another without doing it from scratch. Inc. described the alleged tactic as a "distillation attack" and explained why it is considered dangerous, while WION dubbed the broader phenomenon "AI cannibalism."
According to The Hans India, Anthropic laid out the allegations in a letter to US senators, saying the operators were linked to Alibaba. That outlet places the dispute against the backdrop of US restrictions affecting access to advanced systems it referred to as Mythos and Fable 5. Anadolu Agency reported the core claim as Alibaba illicitly accessing the Claude model.
At the time of these reports, the items summarized here present Anthropic's allegations; they do not include a detailed response from Alibaba confirming or denying the claims.
Why it matters: the case highlights how the most valuable asset in AI — the behavior of a leading model — can potentially be copied at scale, turning a commercial rivalry into a national-security and policy fight playing out in front of US lawmakers.