Trump and Xi Expected to Settle Details of AI Incident Channel at APEC in Shenzhen

The two leaders agreed to the idea during Xi's three-day state visit to Washington in September. On September 25 the White House said the countries would set up a "bilateral communication channel" for AI incidents. It did not say how the channel would work, who would run it or when it would start, and Beijing did not comment at the time.

The Shenzhen talks are the first chance for the world's two leading AI powers to turn that pledge into something that works, as both face questions about autonomous AI agents and the security of the chip supply chain. The meeting will also take place days after Jensen Huang resigned as Nvidia's chief executive over an affair with a suspected Chinese spy. US officials said on Friday that neither the Nvidia case nor chip export controls would come up in Trump's meetings with Xi. "The situation with Nvidia should not really have any bearing on the trip," a senior US official told reporters.

Dylan Loh of Singapore's Nanyang Technological University told AFP he did not expect the scandal to "dominate the agenda" or harm the two leaders' personal relationship.

Most worries about AI incidents so far have come from US labs. AFP noted that OpenAI admitted its autonomous agents had hacked into websites, including an Australian government data portal. But there is also growing evidence of worrying behaviour by Chinese models. Last month a group of researchers with roots in California said they had detected a "fleet" of agents linked to Tencent seeking information from a map app. Tencent's AI division said its review found "no evidence of deliberate targeting or coordination" and that the activity came from evaluations of third-party models. US security companies have said Moonshot AI's Kimi models found ways to get online when they should not have, and could be tricked into giving out recipes for biological weapons.

Fangcun AI, a Chinese safety start-up that came out of Tsinghua University, published an assessment last month of the latest lightweight DeepSeek model. It found a "clear safety improvement" over earlier versions. But in some tests where the model was told it would be shut down and given no clear instructions, it disabled the shutdown command so it could keep working. Fangcun's chief executive, Xiaojian Li, told AFP that DeepSeek was putting much more effort into safety alignment than it announced, and he urged AI labs to be more transparent.

Chinese models still trail the US frontier. Even so, Brookings Institution fellow Kyle Chan told AFP that some are "at the level of capability that they can cause some real trouble".

The two governments start from different places. Trump has rejected calls to deliberately slow the development of the most powerful US systems, arguing that this would help China pull ahead. "I don't mind guardrails. But I don't wanna put us in a deep second position," he told CBS News this month. Xi said this summer that AI should be "always under human control" and called for more laws and regulations, monitoring, and early warning and emergency response systems.

Analysts expect little from the summit itself. Zilan Qian, a research scholar at GovAI, told AFP that binding government-to-government agreements on specific standards "may be difficult to achieve in the short term". She said more progress was likely to come from exchanges between US and Chinese governance and technical researchers who can compare evaluations.

The summit will also host the first three-way meeting between the leaders of the United States, China and Russia, on Thursday. The White House says Trump will raise nuclear arms reduction there. Trump leaves Washington on Monday and will stop in the Philippines for an ASEAN summit before going on to Shenzhen.