US and China Must Cooperate on AI Safety, Experts Warn at Beijing Conference
Key Takeaways
- ▸Frontier AI cybersecurity and systemic risks are global in nature; both U.S. and China recognize that reckless AI development threatens both nations equally
- ▸International cooperation on AI safety can coexist with economic competition; experts point to successful Cold War-era nuclear collaboration as a proven model
- ▸Shared technical standards and safety principles on agentic AI, code vulnerabilities, and automated attacks would benefit both nations while maintaining security
Summary
At a major AI conference in Beijing organized by the Beijing Academy of Artificial Intelligence, industry experts and researchers from around the world gathered to discuss frontier AI challenges, including cybersecurity risks and systemic dangers posed by increasingly capable AI models. The conference featured renowned figures including Whitfield Diffie (co-inventor of public-key cryptography) and Turing Award winner Andrew Barto. A central message emerged: the fierce US-China AI rivalry poses a serious risk to global safety, and the world's two dominant AI powers must prioritize cooperation on AI safety threats—even as they compete economically.
The U.S. government has been tightening restrictions on China's AI development through export controls on advanced chips and chipmaking equipment. Most recently, U.S. authorities ordered Anthropic to prevent foreign nationals from accessing its most advanced models, Mythos and Fable 5, over national security concerns; Anthropic subsequently revoked access for all users. However, experts at the conference, including MIT computer scientist Stephen Casper, argued that international collaboration on AI risks is essential and could outweigh national security concerns. Casper drew a parallel to Cold War-era nuclear cooperation between the U.S. and Soviet Union—nations that intensely competed yet recognized the mutual benefit of jointly addressing existential threats.
The conference highlighted shared cybersecurity challenges that transcend national borders: AI-generated code vulnerabilities, novel attack vectors enabled by agentic AI, and automated social engineering attacks. Researchers like Lin Yun from Shanghai Jiao Tong University emphasized that finding "areas where sharing can reduce systemic risk without exposing sensitive operational details" could be key to progress. As AI becomes more capable and intertwined with critical infrastructure, the urgency of establishing shared safety principles and technical standards becomes clearer to experts on both sides.
Editorial Opinion
The most compelling insight from this conference is that AI safety is fundamentally a shared challenge that transcends geopolitics. While export controls and investment restrictions reflect legitimate national security concerns, they risk pushing AI development into isolated silos where safety could be an afterthought. The U.S. and China have the expertise and resources to co-develop safety frameworks—and their cooperation could set a template for global AI governance that many other nations would follow.
