The US and China can barely agree on trade these days. But they just sat down to discuss AI safety, and one model is the reason.
For years, Washington has tried to get Beijing to the table on AI risks. In 2024, a large American delegation traveled to Geneva for what they hoped would be productive talks. The Chinese side dismissed their concerns about dangerous AI as "a bit abstract, and not really as relevant" and quickly steered the conversation toward US export controls. The delegation came home with nothing.
Then Anthropic's Mythos came along.
As we covered Thursday, the model's cybersecurity capabilities were alarming enough to push the White House toward mandatory pre-release reviews for frontier models. But the international consequences may be the bigger story. Jake Sullivan, who served as national security advisor under Biden and initiated early AI talks with China, told the Los Angeles Times that the Chinese regarded American concerns as "not really as relevant" to their own AI focus. He said he urged the incoming Trump team to continue the dialogue.
The sticking point was always Beijing's suspicion. Sullivan said Xi Jinping and Wang Yi privately supported AI cooperation at "a conceptual level," but their urgency was "considerably lower" than Washington's. China, he explained, tends to view any American diplomatic initiative involving restrictions on a capability "as being a trap."
Mythos shifted that. Daniel Remler, who led AI policy at the State Department and took part in the Geneva talks, cast doubt on Chinese claims of disinterest in AGI and ignorance of its risks. They'd already been open about how compute-constrained they were at the frontier. Now an American lab had built a model that could autonomously find and exploit decades-old vulnerabilities in every major operating system and web browser, and their own models couldn't come close. The New York Times reported that China sought direct access to Mythos and was turned down.
That combination of capability shock and exclusion did what years of diplomacy couldn't. China issued its own AI agent guidelines on May 8, four days before the latest summit, signaling that Beijing wants to show up with its own regulatory framework rather than negotiate on Washington's terms alone.
The fear extends well beyond foreign ministries. Treasury Secretary Scott Bessent convened Wall Street CEOs alongside Fed Chair Jerome Powell specifically to address the financial system's exposure to Mythos-level cyber threats. Anthony Grieco, Cisco's SVP and Chief Security and Trust Officer, said that "AI capabilities have crossed a threshold that fundamentally changes the urgency required to protect critical infrastructure, and there is no going back." Xue Lan, a senior Chinese government adviser, captured the shared anxiety more simply: "If one country is not safe, then none of us are safe."
The last time the US and China agreed on tech guardrails, from WTO accession to Obama-era cyber norms, Washington spent years regretting the terms. Those deals were negotiated when both sides thought they had time. Nobody thinks that anymore. Fear is the most honest foundation for a negotiation, and right now both superpowers are genuinely afraid of the same thing. Whether that produces lasting cooperation or just another round of summits and press conferences depends on something neither government controls: how fast the next Mythos-level model shows up.
