Anthropic says Alibaba used 25,000 accounts to train Qwen on Claude
Anthropic told the US Senate that operators linked to Alibaba used nearly 25,000 fake accounts for more than 28.8 million exchanges with Claude between April 22 and June 5, a period of about six weeks. Anthropic says they did not hack its systems; they accessed the API like ordinary customers but sent requests on an industrial scale. The alleged goal was to collect Claude's coding and multi-step decision-making abilities and use the answers to train Qwen.
Anthropic called it the largest it has ever found, larger than its cases involving , Moonshot, and MiniMax combined. It is not clearly established that this method is illegal under current law. Anthropic therefore wrote to Congress seeking policy action rather than announcing a lawsuit, and connected the issue to restrictions on exporting Fable 5.
The figures and claimed purpose come from Anthropic and have not been independently verified.
Key points
- Anthropic alleges that nearly 25,000 fake accounts produced more than 28.8 million Claude exchanges.
- The activity allegedly ran for about six weeks, from April 22 through June 5.
- The operators reportedly used the API as customers rather than hacking Anthropic's internal systems.
- Anthropic says Claude's answers were collected to improve Qwen's coding and multi-step decision-making abilities.
- The legal status of large-scale remains unclear, and Anthropic's allegations have not been independently verified.