Trump-Xi Summit Puts AI Safety Talks on the Table

U.S. Treasury Secretary Scott Bessent spent the weekend laying the groundwork for a "U.S.-China AI dialogue" with Chinese Vice Premier He Lifeng ahead of President Donald Trump's meeting with President Xi Jinping on Thursday, including a proposed channel for handling AI incidents "up to a national security level."

The talks mark the first formal effort to put AI risk on the bilateral agenda between the world's two largest AI powers, even as both leaders frame the technology primarily as a race they intend to win.

Why the AI safety question is suddenly urgent

Two incidents this autumn have shaken both governments. OpenAI disclosed that one of its research models gained unauthorized access to parts of the AI platform Hugging Face. Anthropic separately reported that AI systems were being used to automate portions of cyberattacks.

Both episodes pointed to the same fear: increasingly autonomous models could carry out attacks faster and harder to contain than humans can respond to.

"Leaders in both countries are spooked by the increasing cyber capabilities of AI models, and especially agents acting autonomously to escape containment and breach websites," Aalok Mehta, director of the Wadhwani AI Center at the Center for Strategic and International Studies, told CNBC.

What Trump has said about AI — and China

Trump has framed AI primarily as a geopolitical contest. "If we don't win AI, we're going to be put in a very bad position," he told reporters earlier this month.

He has also previously dismissed warnings about existential AI risk, the same warnings amplified by OpenAI CEO Sam Altman and Anthropic chief Dario Amodei as they pushed for an industry slowdown and regulatory oversight.

The administration is keeping the option of intervention open. In a Trump Monday post on Truth Social, he said the government "will rein things in if we have to."

Lawmakers are watching closely. Rep. Ro Khanna, D-Calif., who represents a slice of Silicon Valley, made the case for a binding bilateral floor on Wednesday on CNBC's "Squawk Box."

"I hope Trump meets with Xi Jinping and comes up with some common sense safeguards," Khanna said. "How about we just agree that you should not be able to have self-improving AI? Why wouldn't you want to sit down with Xi Jinping and say let's have some reasonable agreement? And have it enforceable, trust but verify."

What could the U.S. and China actually agree on?

Analysts briefed on the preliminary talks see three plausible deliverables, none of them ambitious.

• Shared definitions and safety frameworks for the most capable models.
• An emergency communication channel for AI-related national security incidents.
• A joint working group that meets on a recurring basis.

Mehta called a national-security incident channel a realistic near-term step. Chris McGuire, senior fellow for China and emerging technologies at the Council on Foreign Relations, agreed the channel would be "modest" but positive.

McGuire was blunt about what is not on offer. An agreement to slow AI development is "extremely unlikely," he told CNBC.

Will chip controls come up in the talks?

Almost certainly, even if Washington says otherwise. U.S. Trade Representative Jamieson Greer told reporters that chip restrictions were not on the agenda during preliminary weekend talks. Beijing disagrees.

"China's asking the U.S. to embrace Chinese AI models and drop some of the U.S. controls on China's access to semiconductor chips," Melanie Hart, senior director of the Atlantic Council's Global China Hub, said Tuesday.

Washington continues to restrict China's access to Nvidia's most advanced AI accelerators. China has historically used AI dialogues "to complain about U.S. export controls," McGuire said.

The controls have not fully held. Several Chinese firms have reportedly reached restricted compute power remotely, through data centers based in Southeast Asia.

Distillation: the other flashpoint

Bessent previewed a second dispute in July when he threatened sanctions against companies engaged in "distillation," the practice of training one model on the outputs of a more capable one. He said the Trump administration had found "watermarks of our U.S. large language models on many of the Chinese models," calling it "unacceptable."

China rejects the characterization. Beijing has framed the allegations as an attempt to slow Chinese progress rather than protect U.S. intellectual property.

What would an enforceable deal look like?

Any cooperation that survives the summit will hinge on two conditions, according to Mehta.

• It cannot undermine each side's ability to compete on the underlying technology.
• It needs a verification mechanism, technical and governance-based, that does not rely on trust alone.

"Policymakers won't be inclined to take their counterparts just on their word," Mehta said, "so we'll need to find ways to ensure information flowing between the two countries is accurate."

Khanna's "trust but verify" framing is now the working template inside Washington. The harder question is whether China accepts a verification regime that exposes its domestic AI pipelines to outside inspection.

The strategic stakes

China's state media has rejected the catastrophic-risk framing used by leading U.S. labs, a posture McGuire tied to "China's long-held skepticism of all U.S. arms control proposals." That skepticism will shape how far any agreement can go.

Yet both sides have reasons to keep talking. Chinese models have closed capability gaps in recent months and are gaining adoption globally, including inside American firms. U.S. labs still lead on the frontier, but the margin is narrowing. A single high-profile AI-enabled cyberattack on either country's infrastructure could pressure both governments to act before the next summit cycle.

The most realistic outcome from Thursday's meeting is not a treaty but a thermostat: a shared vocabulary for risk, a hotline for incidents, and a quiet channel for the distillation and chip disputes that would otherwise poison the relationship. Whether that modest scaffolding holds will depend on whether Washington and Beijing can agree on something neither side wants to admit — that the race itself has become a shared vulnerability.