When President Trump arrives in Beijing this week for a state visit with Chinese President Xi Jinping, the agenda will include something neither administration expected to be discussing: artificial intelligence safety. In a significant reversal from the Trump administration's earlier laissez-faire posture on AI governance, quiet discussions have taken place ahead of the summit to explore reviving an emergency communication channel between Washington and Beijing on AI -- a diplomatic initiative that stalled under Biden and was abandoned entirely in the early Trump years.

The catalyst, according to officials briefed on the discussions, is Anthropic's Mythos model. Released last month, Mythos has been assessed by both US and Chinese government analysts as representing a qualitative leap in AI capability -- one with direct implications for national security. The model's ability to autonomously identify and exploit vulnerabilities in digital communication systems, financial infrastructure, and government databases has alarmed officials in both capitals.

The Beijing summit will be the first time AI safety features prominently on a US-China leaders agenda. Image: SUPERBASH_
The Beijing summit will be the first time AI safety features prominently on a US-China leaders agenda. Image: SUPERBASH_

From Woodside to Beijing

The history of US-China AI diplomacy is short and largely unsuccessful. In 2023, Biden and Xi held their first high-level talks on AI risks in Woodside, California. The following year, a US delegation travelled to Switzerland for formal talks, only to find the Chinese contingent dismissive of American concerns about frontier AI risk. Beijing viewed the discussion as a pretext for export controls, not a genuine safety dialogue.

The one concrete outcome of that period was a November 2024 accord in Peru, where both sides agreed to keep AI out of the command and control of nuclear weapons. Jake Sullivan, Biden's national security advisor, described it as a 'breaking of the seal' -- proof that some agreement was possible. He urged the incoming Trump team to continue the dialogue. They did not.

"That's all changed in the past few weeks. A Trump administration once eager to gun for technological supremacy is now, for the first time, reckoning with the power AI could unleash if left unchecked."

— Jake Sullivan, former US National Security Advisor

The Mythos Factor

What changed the calculus was Mythos. The model's capabilities have been described by industry insiders and government analysts as those of an unprecedented cyberweapon -- one capable of infiltrating and exploiting digital systems at a scale and speed that no human team could match. Mozilla's security research team identified hundreds of novel attack vectors in Mythos within weeks of its release. The model can reportedly generate working exploits for previously unknown vulnerabilities faster than most organisations can patch known ones.

For the Trump administration, which had positioned AI supremacy as a core national security objective, Mythos represented a problem: the technology it was racing to develop had become powerful enough to threaten the systems it was meant to protect. For Beijing, which had long argued that US AI development was reckless, Mythos was evidence that the race dynamic itself was dangerous.

What an AI Hotline Would Look Like

The proposed emergency channel is modelled on existing communication mechanisms between the two powers -- the nuclear hotline established after the Cuban Missile Crisis, and the military-to-military communication channels used to prevent accidental escalation in the Taiwan Strait. A senior administration official told reporters on Sunday that the White House was looking to create a channel 'like others that they have in many areas that have intense focus with the US and China,' adding that its formality was 'yet to be determined.'

Experts are cautious about what such a channel could achieve. Aalok Mehta of the Wadhwani AI Center at CSIS noted that there is 'almost no support from US policymakers to engage in formal discussions on AI governance with China,' given the prevailing view that the AI race is winner-takes-all. Any agreement would face deep scepticism in Congress, where bipartisan consensus holds that China cannot be trusted to honour commitments on technology.

The more immediate question is whether the summit produces any announcement at all. The discussions are described as exploratory, and the gap between the two sides' strategic interests remains wide. The US is racing toward AGI; China is focused on AI diffusion into industry. The US wants guardrails on frontier models; China wants the US to lift chip export controls. Finding common ground in that context is a diplomatic challenge of the first order -- but the fact that both sides are talking at all represents a shift that few predicted.