Anthropic accuses Alibaba of largest distillation attack on Claude, 28.8M model queries via 25K fake accounts

Anthropic sent a letter to the U.S. Senate Committee on Banking, Housing, and Urban Affairs accusing Alibaba of orchestrating the largest known distillation attack on Claude to date. Between April 22 and June 5, 2026, operators affiliated with Alibaba and its Qwen AI lab conducted 28.8 million exchanges with Anthropic's models using approximately 25,000 fraudulent accounts, circumventing Anthropic's geographic access restrictions and terms of service.

Distillation is a machine learning technique where a smaller, less-capable model is trained on outputs from a stronger model. While legitimate in house, illicit distillation allows competitors to acquire frontier capabilities in a fraction of the time and cost required to develop independently. Anthropic noted the campaign mirrors similar attacks flagged in February by DeepSeek, Moonshot AI, and MiniMax—three other Chinese labs that collectively generated over 16 million exchanges via 24,000 fraudulent accounts. Alibaba did not immediately respond to requests for comment; shares fell 3% on the news.

For practitioners: distillation attacks represent a new threat vector that scales as AI APIs become more valuable. The targeting of agentic reasoning, tool use, and coding capabilities suggests systematic IP theft aimed at narrowing the capability gap between frontier and localized models. Anthropic flagged this requires coordinated action between government and industry. The disclosure arrives amid broader U.S. export controls: Anthropic recently received orders to suspend its latest Claude models (Fable 5, Mythos 5) from foreign nationals, now under negotiation with the Trump administration.