U.S. Treasury Secretary Scott Bessent confirmed Thursday that Washington and Beijing have begun establishing a formal protocol for AI safety best practices during President Donald Trump's two-day summit with Chinese President Xi Jinping in Beijing. The discussions mark the first high-level bilateral framework aimed at preventing non-state actors from accessing frontier AI models, according to remarks Bessent made to CNBC's Joe Kernen on the sidelines of the diplomatic meetings.

'The Two AI Superpowers Are Gonna Start Talking'

Bessent was blunt about why the U.S. is now positioned to negotiate with China on AI governance. "The reason we are able to have wholesome discussions with the Chinese on AI is because we are in the lead," he said. "I do not think we would be having the same discussions if they were this far ahead of us." The Treasury Secretary outlined the protocol's primary objective: ensuring that powerful AI systems don't fall into the hands of terrorist groups, ransomware operators, or other malicious non-state entities. Bessent also indicated that upcoming large language model releases from Google's Gemini and OpenAI would represent a "big step-function jump" in capabilities.

Anthropic's Mythos Model Raises Alarms

The urgency behind these talks appears amplified by concerns over Anthropic's recently announced Mythos AI model, which reportedly possesses significant cyberattack capabilities. Washington and allied governments have expressed alarm at the model's potential dual-use risks. Anthropic has stated it will initially restrict Mythos to select business partners rather than a public release—a containment strategy that underscores how seriously the industry is taking proliferation concerns. The timing of these bilateral talks suggests policymakers want guardrails in place before more frontier models reach similar capability thresholds.

Nvidia Chip Restrictions Remain Contentious

The semiconductor dimension of U.S.-China AI competition surfaced prominently during the summit, with Nvidia CEO Jensen Huang joining Trump's delegation as a late addition to the official party. Washington has pursued aggressive export controls on advanced AI chips—primarily Nvidia's hardware—to limit China's development trajectory. However, Bessent acknowledged "a lot of back and forth" regarding Reuters reporting that the U.S. had cleared sales of Nvidia H200 chips to several major Chinese technology firms. The mixed signals highlight the tension between maintaining technological advantage and preserving commercial ties with a critical trading partner.

Taiwan Dominates Geopolitical Agenda

Beyond AI governance, Xi used Thursday's meeting to deliver a stark warning about Taiwan, calling it "the most important issue" in bilateral relations and cautioning that mishandling could push both nations toward "conflict." Beijing claims the democratically self-ruled island as part of its territory. Bessent indicated Trump would address Taiwan concerns more fully "in the coming days," describing the President's approach as reflecting deep understanding of the sensitivities involved.

Key Takeaways

  • US and China establish first bilateral AI safety protocol targeting non-state actor access to frontier models
  • Anthropic's Mythos model with cyberattack capabilities cited as driver for urgent governance talks
  • Nvidia CEO Jensen Huang joins Trump's Beijing delegation amid chip export policy ambiguity
  • Xi warns Taiwan mishandling could trigger superpower conflict during Thursday's summit

The Bottom Line

Let's be real: this is geopolitics dressed up as AI safety theater. Washington's "we're in the lead" confidence conveniently aligns with wanting a seat at the table for writing the rules. But here's what matters to builders and hackers watching from the sidelines—these talks will shape export controls, model access policies, and eventually the open-source landscape. The Mythos situation proves frontier labs are already self-regulating because they have to. Whatever protocol emerges from Beijing will define constraints you won't escape.