The covert U.S.-China battle to make chatbots leak their secrets - The Washington Post

Anthropic has alleged in a letter to U.S. senators that Chinese tech firm Alibaba's Qwen AI team used approximately 25,000 fraudulent accounts to generate more than 28.8 million exchanges with its Claude model. The letter states that these exchanges were used to 'distill' knowledge from Claude to improve Alibaba's own AI technology. Anthropic also disclosed that it had deployed tracking code specifically designed to detect such activities by Chinese firms.

The 25,000 accounts and 28.8 million exchanges represent a significant scale of alleged model theft, though the letter did not specify the time period over which these interactions occurred. Anthropic has not publicly released the technical details of the detection mechanism, citing competitive sensitivity.

The allegations come amid escalating U.S.-China tensions over AI technology, with both governments imposing restrictions on AI chip exports and model transfers. If confirmed, the incident would represent one of the largest known cases of model distillation via API access. The senators who received the letter have not yet publicly commented, but the matter is likely to fuel further calls for stricter controls on AI model access by foreign entities.