Anthropic has told U.S. senators that Alibaba ran the largest known distillation attack on an American AI model, harvesting nearly 29 million exchanges with Claude through 25,000 fake accounts to train Qwen, and the resulting capabilities are already showing up in benchmark results.
The letter went to Senate Banking Committee Chair Tim Scott and Ranking Member Elizabeth Warren on June 10. Anthropic alleged that operators affiliated with Alibaba's Qwen lab opened roughly 25,000 fraudulent accounts and used them to generate 28.8 million exchanges with Claude between April 22 and June 5, 2026, specifically targeting the model's agentic reasoning, software engineering, and long-horizon task capabilities. The goal, as Anthropic characterized it, was illicit capability extraction: train a less capable model on outputs from a more powerful one and close the gap without doing the underlying research. As CNBC reported, Anthropic described Alibaba's conduct as
Also read: California built a real-time dashboard to track whether AI is actually killing jobs • China's AI and chip companies are flooding domestic markets with IPOs as the door to New York stays shut • A single Broadcom guidance miss sent South Korea's stock market into circuit-breaker territory and exposed how much the AI trade depends on perfection