Just behind a week ago I attended a enormous artificial intelligence conference in Zhongguancun, a bustling high-tech district in Beijing.
It was filled with fascinating sessions on topics ranging from recursive self-improvement – the idea that models can modify their own code and evolve indefinitely – to humanoid robots. It featured several computer science legends, including Whitfield Diffie, co-inventor of public key cryptography, and Andrew Barto, who won a Turing Award with Opulent Sutton for his pioneering work on reinforcement learning.
But I came away with one conclusion above all: The United States and China should put aside their fierce artificial intelligence rivalry.
Frontier AI’s cyber and systemic threats are too stern to ignore, and increasingly capable agent models could soon cause chaos if the world’s AI superpowers cannot cooperate. “Artificial intelligence is a global technology that brings global benefits, global harms, and a constant tendency to ultimately spread new opportunities.” Stefan Kacperan MIT computer scientist who spoke at the conference via video told me later.
Until now, the United States has largely viewed China’s advances in artificial intelligence as a threat to its economy and national security. Washington has imposed strict restrictions on chips and chip-making equipment to hamper the development of powerful artificial intelligence in the country. Recently, the US government ordered Anthropic to prevent foreigners from accessing its most powerful models, Mythos and Fable 5, for national security reasons. In response, Anthropic revoked everyone’s access. As WIRED previously revealed, the company of particular concern was a South Korean telecommunications giant with alleged ties to China.
However, the conference organized by the Beijing Academy of Artificial Intelligence reinforced the belief that both the United States and China stand to lose if artificial intelligence is developed too quickly and recklessly. As AI becomes more powerful, more agentic and more interconnected with everyday life, the risk of it being used to launch cyberattacks or cause catastrophic failures will only boost. Since two of the world’s dominant AI powers are responsible for the most advanced models, it seems that cooperation between them will be crucial.
Kacper pointed tests demonstrating that the benefits of international cooperation on AI threats outweigh any national security risks arising from cooperation. He compared the current situation to one in which the United States and the Soviet Union were forced to cooperate in the fight against nuclear threats even as they tried to stockpile each other’s supplies.
“The one thing that almost everyone in AI can agree on right now is that AI doesn’t need a Chernobyl moment,” Casper said.
The one-day session highlighted the universality of the cyber challenges posed by more advanced artificial intelligence. This includes fresh types of vulnerabilities in AI-generated code, novel ways of attacking systems made possible by the apply of agentic tools, and automated methods of conducting social engineering attacks.
After another session, I talked to Lin Yun, a professor at Shanghai Jia Tong University, who he does a great job on artificial intelligence and computer security. Yun told me he expects hackers to gain the upper hand in the near future, but fresh countermeasures, including novel applications of artificial intelligence, should tip the scales back toward defense over time.
Yun said that even if international cooperation is complicated by competition, it should remain a priority. “If different countries understand the risks in a similar way, it is easier to develop common safety rules and technical standards,” he told me. “The key is to find areas where sharing information can reduce systemic risk without revealing sensitive operational details.”
Perhaps the most pressing question for both nations is how to balance openness with risk. Open-scale models have become central to research and innovation, and Chinese models are proving popular in the US. However, as these models develop, it will become an increasing challenge to ensure that they do not aid hackers identify vulnerabilities and cannot be used as cyber weapons.