The US and China are about to sit down for another round of AI safety talks — and this time, a rogue model from Anthropic and a shift in Chinese chip strategy are driving the conversation.
The direct catalyst is Anthropic's frontier model, Mythos, which can autonomously hack into government databases, banks, and hospital networks. US officials call it an "unprecedented cyber weapon." Even though Mythos was built in America, it proves this kind of destruction is technically feasible. If a rival or hacker group gets similar capabilities, neither the US nor China would be safe. At the same time, DeepSeek announced for the first time that its new model has been adapted to run on Huawei chips — a sign that Chinese AI is starting to wean itself off dependence on US chip giant Nvidia. That gives Beijing a stronger hand at the bargaining table. Senior US officials confirm the White House wants to use the window of an upcoming meeting between the two countries' leaders to open a communication channel.
The Biden administration had already pushed for US-China AI dialogue twice. In 2024, during the first round of technical talks in Geneva, the US side worried about "AI losing control," but according to four participants, Beijing treated that as an academic hypothetical and instead pressed for an end to chip export restrictions. The talks went off the rails. In November of the same year, the two sides reached a limited agreement in Peru, pledging not to let AI control nuclear launches. Then-National Security Advisor Jake Sullivan called the Peru deal a "break of the seal," proving the US and China could produce substantive AI agreements. He strongly recommended the incoming team continue the talks, but they didn't take his advice. It wasn't until Mythos appeared that Washington's attitude really shifted. US Treasury Secretary Bessent, after hearing security briefings from multiple banks, proposed federal review of future new models, and the White House finally changed course.
But a deeper obstacle is that the two countries are running completely different races. Silicon Valley giants are sprinting toward artificial general intelligence (AGI) — AI that matches human-level general intelligence — believing that whoever builds AGI first will achieve irreversible technological dominance. China isn't chasing the smartest AI; it's embedding existing AI into factories, hospitals, autonomous vehicles, and government systems to generate immediate economic value. "China doesn't see a single race, but multiple parallel races," says CSIS senior adviser Kennedy. When the two sides sit down, the US wants to talk about "what happens if superintelligence goes rogue," and China wants to talk about "why won't you let me buy chips."
Stanford's latest report says the performance gap between US and Chinese models has "essentially closed," but the US still holds advantages: US tech companies spend 10 times more on AI than their Chinese counterparts, and they run 10 times as many data centers (the physical infrastructure for AI computing). Some industry experts suggest breaking off specific risks one at a time, like the Peru nuclear agreement. But Kyle Chan of the Brookings Institution warns that mutual distrust is fueling a "race to the bottom" in security — both sides are cutting safety investments to move faster. Sullivan says the Chinese leadership agrees in principle that cooperation is necessary, but their sense of urgency is far lower than Washington's. "Their urgency has increased, though," he adds.