A line to pick up when something breaks
The United States will propose a notification process with China for incidents involving artificial intelligence, Treasury Secretary Scott Bessent said in an interview with The Axios Show published on Saturday. He said he expects Beijing to agree to one.
Bessent said he and Chinese Vice Premier He Lifeng discussed AI safety ahead of President Xi Jinping’s state visit in September, and that a proposal for an incident communication channel was part of those talks. The risks the two sides covered, he told Axios, included uncontrolled AI agents and the use of the technology by non-state actors for cyber or biological attacks.
The idea is narrow by design. It is not a safety treaty, not a shared standard and not an inspection regime. It is a line one government picks up when a system has misbehaved badly enough that the other needs to know. That is a modest ask in a relationship where the Trump administration has refused formal AI safety cooperation with Beijing.
Why Bessent thinks China would say yes
His argument rests on a claim about how Beijing’s own thinking has moved. “I think that they didn’t realize how powerful their open source models are,” Bessent said, referring to Chinese models that anyone can download and adapt.
He pointed to an incident involving Kimi, the model family built by Moonshot AI, in which he said the model sent sensitive Chinese data to Anthropic. That, in his telling, was one of the developments that “made the Chinese realize that this is a very powerful technology.”
“A lot of these Chinese models are very powerful — not as powerful as the U.S. models — but you get something that’s 80%, 90% as powerful, without guardrails that can be imported anywhere in the world,” Bessent said.
Beijing has its own anxieties about the technology, though not the same ones. China’s spy chief warned in September that AI could be used for information and “cognitive” warfare against the government — a framing that locates the threat in the Communist Party’s grip on information rather than in public safety.
Voluntary reviews, and no liability shield
At home, Bessent described a light-touch settlement that leaves responsibility with the companies. Government reviews of frontier models are voluntary, he said, though the administration reserves the right to intervene if a laboratory presses ahead despite safety concerns.
“It is the people at the labs who have to accept responsibility, and I agree with that,” he said. “I think the labs have come around to that way of thinking, too.”
He was sharper about the executives who have spent recent weeks warning publicly that they may lose control of their own systems. Asked about those warnings, Bessent replied: “Well, then they should slow down.” In the same interview he described the pattern as “alarmism without solutions” and said that is not leadership.
What he says he wants instead is a shift of attention from capability to defence. “We are doing this hard scramble to the frontier for the models,” he told Axios. “We really need to start worrying more about resilience and defense, too.”
What to watch
Bessent gave no date for putting the proposal to Beijing, named no agency to run the channel, and did not say what would count as a reportable incident. Those are the details that separate a working mechanism from an announcement. The next test is whether the channel turns up in any readout of the US-China talks that follow Xi’s visit.
