Alibaba's 28.8m exchanges reveal China's AI extraction strategy

Anthropic disclosed the largest known attack against its Claude AI model, accusing Alibaba and its AI lab Qwen of systematically extracting capabilities between April and June 2026.

The AI  company disclosed the largest known attack against its Claude AI model, accusing Alibaba and its AI lab Qwen of systematically extracting capabilities between April and June 2026.

The Alibaba attack is on a totally different level compared to those executed by previous Chinese AI labs. Using over 25,000 fake accounts, the hackers were able to create 28.8 million exchanges with Claude, which surpasses the previously used 150,000 exchanges campaign by DeepSeek revealed in February and the combined total of 16.4 million exchanges by Moonshot AI and MiniMax.

According to Anthropic, the attack is known as "distillation," which refers to training weak models using the outputs produced by more powerful ones. Rather than coming up with their own AI systems from scratch, Chinese labs take the responses produced by Claude and come up with less expensive and quicker AI systems.

The announcement of three independent Chinese labs (DeepSeek, Moonshot, and MiniMax) in February indicated sporadic cases.

Anthropic informed the senators that the attacks are “growing in intensity and sophistication,” which indicates involvement or coordination at the state level.

Two days after Anthropic’s letter was sent to the Senate on June 10, the Department of Commerce prohibited the Mythos and Fable models of Anthropic from being deployed worldwide due to concerns over military use in China.