The U.S. Treasury Secretary revealed that the United States and China are discussing the establishment of an emergency notification mechanism for AI accidents

U.S. Treasury Secretary Scott Bessent revealed that the United States and China are discussing the establishment of an emergency notification mechanism for artificial intelligence accidents. (AP file photo)

U.S. Treasury Secretary Scott Bessent said that the United States is preparing to establish an artificial intelligence (AI) accident notification mechanism with China. If major problems occur in related systems, both parties can notify each other. He also used the data transfer incident involving the Chinese AI model "Kimi" as an example to illustrate the security risks that open source AI may bring.

According to the US news website "Axios", Bessant discussed the US-China AI competition in an interview with "Axios Show" on the 3rd, revealing that he recently discussed AI safety issues with Chinese Vice Premier He Lifeng, including the establishment of an emergency communication channel for major bilateral AI incidents. Bessent said that the United States cannot lose its current leading position in the field of AI.

Please read on...

The Kimi incident mentioned by Bessant originated from the threat intelligence report released by the American AI company Anthropic in September. The report alleged that the Chinese AI startup "Moonshot AI" used about 5,380 fake accounts to secretly forward some users' question requests that were originally handled by Kimi to Anthropic's Claude model for calculation. Anthropic said that during its investigation, it found that nearly 300,000 requests were forwarded in 10 days.

Anthropic further pointed out that one of the users assessed to be "likely" associated with the People's Liberation Army of China used the model to analyze image data from hundreds of monitors in Chengdu, including cameras outside Chinese military facilities. These surveillance data also entered Claude through the transfer mechanism.

He once said that the leak of weapons information became the reason for security negotiations

At a hearing of the U.S. House of Representatives Financial Services Committee on September 15, Bessant mentioned that Kimi had a security incident and sent Chinese military weapons programs back to Anthropic. However, this statement is not the same incident as the transfer of surveillance images recorded in the Anthropic report and another case in which personnel associated with the Chinese Academy of Military Sciences used Claude to develop air defense suppression software.

The Anthropic report stated that Dark Side of the Moon used answers generated by Claude to train its own AI model, also known as "distillation," between May and July 2026. During this period, more than 23 million model exchanges were observed.