Anthropic exposes Alibaba's large-scale model distillation attack

In a report recently submitted to the U.S. Senate, Anthropic revealed that Alibaba used 25,000 fake accounts to have 28.8 million conversations with Claude over about six weeks between April and June.

This incident was characterized not as simple data collection but as a so-called 'Distillation Attack' that extracted Claude's Agentic Reasoning and coding capabilities and used them to train its own model, Qwen. Anthropic described this as the largest attack in its history.

The key characteristics are as follows:

• Non-intrusive method: Data was extracted at an industrial scale through normal API usage rather than hacking.
• Legal gray area: Under the current legal system, it is difficult to clearly classify such large-scale distillation as illegal.
• Response approach: Instead of litigation, Anthropic submitted a report to Congress, emphasizing the need for policy responses and regulation.