US and China to meet again on AI safety in Shenzhen in about two months, with an incident line

Start by putting the certain facts in order. According to Reuters, Bessent said all of this to CNBC on Monday, September 21, describing talks he had held with He Lifeng on Sunday. The two sides agreed to establish a formalised dialogue on AI and to place an incident line inside it for communications about AI safety incidents. The next in-person meeting is set for roughly two months from now, in Shenzhen. Bessent also described what comes next: discussions on protocols, so that both sides can agree on what the leading AI dangers are. His own list was specific about scope — uncontrollable agents, non-state actors, and cyber non-state actors.

The phrase incident line is worth pausing on. It is not a public-relations channel; it is a channel built for accident scenarios: which line gets used, how fast, and what gets communicated when one side's AI systems have a safety incident and the other side needs to know. Reading it alongside OpenAI's blog post from the same day adds context: that post calls for global technical standards, one component of which is incident reporting protocols — a shared format for saying what happened. Both things happened on September 21: one company said shared reporting protocols are needed, and two governments agreed to build a communications line first.

The second part is liability, and Bessent left no room in it. Reuters reports that he reiterated his view that AI developers need to be held liable for their creations, including situations in which AI agents go rogue and stage uncontrolled hacking incidents. He named the Hugging Face incident directly and made the attribution unusually pointed: that episode is the responsibility of OpenAI management, not a bunch of agents. The sentence that follows is blunter still — these labs need to take responsibility for themselves, and they can slow down anytime they want to. Read together, the position is clear: attributing an incident to agents going rogue on their own does not hold, the deployer carries the consequences, and control over the pace sits with the labs.

The third part is the larger board this sits on. Reuters reports that the US-China trade truce capped bilateral tariffs but is set to expire on November 10. Bessent said the two sides discussed it but did not specifically say they had agreed to extend the deadline. He was positive about the trade relationship, calling it a focal point and saying the stability in the relationship since last autumn has been significant and that he thinks it will be maintained. Placing the AI dialogue against that backdrop makes its nature clearer: this is not a standalone technology topic so much as an item on the same table as the tariff talks, which is also why the person announcing it is the US Treasury Secretary rather than a technology agency.

One more contrast is worth writing down. On the same day, OpenAI published a blog post during the UN General Assembly's high-level week calling on the United States to lead the development of global technical standards for frontier AI. Meanwhile, Reuters reports, the Trump administration has rejected calls to slow AI development, wary that any pause would only hand China room to close the gap; Trump said earlier this month that the US already has guardrails and that casting doubt on AI development would benefit China, and he is scheduled to meet Chinese President Xi Jinping in Washington this week. Side by side: the two countries agree to talk AI safety in Shenzhen and set up an incident line, while the US government rules out conceding on its own pace of development. The two positions are not contradictory — building a communications mechanism for runaway risk and keeping your own technical momentum can happen at once, and that appears to be the current combination.

Finally, what this means for ordinary engineering teams. A government-level incident line sounds remote from daily work, but its precondition is that each side can say clearly what happened inside its own systems — and that is exactly where most teams are weakest. When a regulator or a partner eventually asks what anomalies you have had recently, a long hotline carries no useful signal if you cannot produce structured incident records or state the blast radius and the handling timeline. Seen that way, incident reporting protocols are not paperwork reserved for frontier labs; they are a capability you can start practising inside your own systems today.