Anthropic accuses Alibaba of stealing Claude's secrets via 28.8M questions

Anthropic has formally accused groups tied to Alibaba and its Qwen AI lab of conducting a massive model-distillation campaign against Claude. In a letter to U.S. lawmakers, Anthropic claims the attackers used nearly 25,000 fake accounts to generate over 28.8 million interactions, deliberately querying Claude about its advanced software engineering and agentic reasoning features. The sheer scale of queries essentially allowed the rival to reverse-engineer Claude's behavior without ever seeing its underlying code or training data. This technique, known as model distillation, exploits the fact that every question an AI answers reveals something about how it works. Anthropic alleges that Alibaba used this information to accelerate its own competing AI development.

The accusations come amid a broader AI arms race where companies guard their models' behavior as valuable intellectual property. Anthropic has previously made similar claims against DeepSeek, Moonshot AI, and MiniMax. OpenAI has also expressed concern about being victimized by this technique. The irony is not lost: AI companies that trained their models on vast public datasets are now arguing those models are proprietary. Anthropic asked lawmakers to act quickly, warning that if frontier models can be easily imitated through mere questioning, the incentive to invest billions in cutting-edge research will evaporate. The next great AI battle, they suggest, may be about preventing copying rather than building smarter models.