Anthropic accuses Alibaba of mass 'distillation' attack to copy Claude

Anthropic has accused Alibaba of running what it calls the largest known attempt to illicitly extract Claude’s capabilities, telling U.S. lawmakers that the Chinese tech giant and its Qwen AI division systematically harvested model outputs through fraudulent accounts. The alleged method is adversarial distillation: training a weaker model on a stronger model’s responses to replicate its behavior while skipping the enormous compute and R&D cost of building comparable capability from scratch.

According to Anthropic’s account, the campaign ran from April 22 to June 5, 2026, generating roughly 28.8 million exchanges across about 25,000 fake accounts, and deliberately concentrated on Claude’s most valuable strengths — software engineering and agentic reasoning. Anthropic laid out the claims in a June 10 letter to Senate Banking Committee Chair Tim Scott and Ranking Member Elizabeth Warren. Alibaba has not publicly responded.

The dispute extends a pattern Anthropic flagged earlier in 2026, when it named DeepSeek and two other Chinese labs over similar extraction efforts; the Alibaba campaign is described as far larger. Anthropic frames distillation as a safety risk as well as an IP one, arguing that models cloned this way often inherit none of the original’s guardrails. The accusation is already feeding policy momentum, with senators preparing defense-bill amendments to sanction Chinese firms that improperly tap U.S. AI outputs — making this as much an export-control and national-security story as a corporate one.