U.S. and China Eye AI Export Restrictions as Models Gain Power

The global AI landscape is shifting from commercial competition to national security confrontation as breakthrough models demonstrate unprecedented autonomous capabilities, prompting both the United States and China to consider restricting access to their most advanced systems.

Three major developments are converging: frontier AI models are achieving dramatic new capabilities, the Trump administration is debating systematic oversight protocols, and both superpowers are exploring export controls on their leading AI technology. These trends were first detailed by Axios based on conversations with AI executives and administration sources.

Why it matters

The emergence of truly autonomous AI agents capable of rewriting software and operating with minimal human oversight is transforming AI from a regulatory afterthought into a matter requiring government intervention. This marks a fundamental shift in how both nations view AI competitiveness—not just as an economic advantage but as a strategic imperative with military and intelligence implications.

Breakthrough models reshape the field

Anthropic's Fable and Mythos models, which faced a three-week restriction in June over security concerns, have established a new benchmark for frontier AI. Engineers report these systems can process multimillion-line codebases autonomously, rebuilding legacy systems, debugging their own code, and conducting self-testing with minimal human supervision.

OpenAI's Sol model, released after what sources describe as a "voluntary" delay for government consultations, represents what early testers call a quantum leap in agentic capability. The system can deploy coordinated sub-agents that collaborate to identify security vulnerabilities and rewrite software at unprecedented speeds.

Elon Musk's SpaceXAI entered the race with Grok 4.5—triple the size of its predecessor—following the company's $60 billion acquisition of Cursor. Musk indicated an even larger model will arrive next month, betting that scale remains a critical factor in AI advancement.

Meanwhile, Chinese startup Z.ai released GLM-5.2 as a free, open-source model performing at the level of America's premium offerings. Founder Jie Tang predicted China will achieve a "Fable-class" model before the first quarter of 2027.

Trump administration weighs systematic oversight

Despite President Trump's preference for light-touch regulation, senior officials are actively debating more prescriptive protocols for AI labs before they release powerful models. The restrictions on Anthropic's Mythos and Fable models demonstrated the administration's willingness to intervene when national security concerns arise.

Commerce Secretary Howard Lutnick's letters to Anthropic signaled that compliance becomes mandatory when security issues emerge. An outside adviser involved in the discussions said "the possibilities are wide open" for new regulatory approaches, including a potential governing body for vetting AI that could involve international participation.

Export controls under consideration

Chinese authorities have recently met with major technology firms to discuss limiting overseas access to advanced Chinese AI models, according to Reuters. In response, U.S. officials are exploring similar restrictions on Chinese access to American models, possibly through export controls, though these conversations remain preliminary.

An administration insider emphasized that AI has already become deeply integrated into both nations' intelligence and military operations, fundamentally altering geopolitical competition and the nature of warfare.

These details were first reported by Axios, with contributions from Zachary Basu and Andrew Kay.