Key facts
- Anthropic alleges Alibaba made 28.8 million Claude AI queries in theft attempt
- Described as the largest-known AI model theft attempt
- US restricts access to AI models Mythos and Fable 5
- Alibaba stock fell to a 16-month low after accusations
- Alleged method: model distillation via mass querying
Anthropic has accused Alibaba of orchestrating the largest-known attempt to steal an AI model, alleging that the Chinese tech conglomerate illicitly accessed its Claude AI system through 28.8 million queries in a bid to distill its capabilities into a competing model. The accusations, detailed in a formal letter from Anthropic, represent a dramatic escalation in tensions between leading Western AI firms and Chinese technology companies.
The technique at the heart of the allegation — model distillation — involves feeding massive volumes of queries to a target AI system and using its responses to train a new, cheaper model that mimics the original's behaviour. Anthropic's letter alleges this is precisely what Alibaba attempted, making it the most significant such effort the company has publicly documented. The US government has separately moved to restrict access to AI models Mythos and Fable 5 in connection with the broader crackdown.
Markets reacted swiftly: Alibaba's stock slid to a 16-month low after the accusations became public, reflecting investor concern about both the legal exposure and the reputational damage to one of China's most prominent technology firms. Alibaba has not publicly confirmed the allegations.
The episode underscores the intensifying battle over AI intellectual property as frontier models become central to national competitiveness. For Anthropic, which counts Claude among its flagship products, protecting the integrity of its models is existential. Google is also reported to be on the verge of losing two more senior AI researchers to Anthropic, signalling that the startup continues to attract top talent even as it navigates this high-stakes dispute.