For all the talk of an AI arms race between the United States and China, a recent gathering of the world's leading artificial intelligence researchers in Beijing pointed to an uncomfortable truth that cut across national lines: the people building the most powerful systems on both sides are genuinely worried about where this is heading, and many believe the two rivals can no longer afford to treat safety as a zero-sum contest.
The conference, organized by the Beijing Academy of Artificial Intelligence, ranged across some of the most provocative questions in the field, from recursive self-improvement, the notion that models could rewrite their own code and keep advancing without limit, to humanoid robots. It also drew genuine pioneers of computing, including Whitfield Diffie, co-inventor of public-key cryptography, and Andrew Barto, who shared the Turing Award with Rich Sutton for his foundational work on reinforcement learning.
One message stood out
Above every technical debate, a single conclusion kept resurfacing: the United States and China should set their bitter AI rivalry aside. The cybersecurity and systemic dangers posed by frontier AI are too grave to brush off, and rapidly improving agentic models could soon trigger real chaos unless the world's AI superpowers find a way to cooperate. "AI is a global technology with global benefits, global harms, and a consistent tendency for new capabilities to eventually proliferate," said Stephen Casper, a computer scientist at MIT who addressed the conference by video.
Washington's hard line
So far, the US has largely treated China's AI progress as both an economic and a national security threat. Washington has placed tight curbs on advanced chips and chipmaking equipment to slow the country's push toward powerful AI. More recently, the US government directed Anthropic to block foreign nationals from using its most capable models, Mythos and Fable 5, citing national security. Anthropic responded by cutting off access for everyone. One firm that drew particular scrutiny was a South Korean telecom giant with alleged ties to China.
Why cooperation may be unavoidable
Yet the Beijing gathering reinforced a different logic: both countries stand to lose if AI is built too fast and too carelessly. As these systems grow more powerful, more agentic, and more woven into daily life, the danger that they could be turned into cyberweapons or fail catastrophically only rises. Since the two dominant AI powers are responsible for the most advanced models, working together starts to look less like idealism and more like necessity.
Casper pointed to research suggesting that the gains from international collaboration on AI risks outweigh whatever national security exposure such cooperation creates. He compared the moment to the Cold War, when the US and the Soviet Union were compelled to manage nuclear dangers together even as each tried to outbuild the other. "One thing that almost everyone in AI can agree on right now is that AI doesn't need a Chernobyl moment," Casper said.
The cyber threat is universal
One full-day session drove home how borderless the security challenges raised by more capable AI have become. They include fresh categories of vulnerabilities in AI-generated code, new avenues of attack opened up by agentic tool use, and automated techniques for running social engineering campaigns.
Lin Yun, a professor at Shanghai Jia Tong University whose work focuses on AI and computer security, expects attackers to hold the upper hand in the near term. Over time, though, he believes new defenses, including novel applications of AI itself, should swing the advantage back toward defenders. Even when competition complicates it, Yun argued, cooperation should stay a priority. "If different countries understand the risks in similar ways, it becomes easier to develop shared safety principles and technical standards," he said. "The key is to find areas where sharing can reduce systemic risk without exposing sensitive operational details."
The open-weight dilemma
Perhaps the thorniest question for both nations is how to weigh openness against risk. Open-weight models have become essential to research and innovation, and Chinese models in particular have found a wide audience in the US. But as they grow more capable, it becomes harder to ensure they don't help hackers spot security holes or get repurposed as cyberweapons.
Over the past few years, Chinese firms have led the way in releasing highly capable open-weight models, among them Moonshot's Kimi, Alibaba's Qwen, and Z.ai's GLM. The US has revived its own open-weight effort with models such as Nvidia's Nemotron. The field is now nearing a tipping point where even less powerful open models could become dangerous once their guardrails are stripped away. Z.ai's latest release, GLM 5.2, carries frontier agentic and coding abilities, according to expert analysis, and the next wave of open-weight models may rival Fable or Mythos. This week alone, 360 Security Technologies, a leading Chinese cybersecurity company, said it had built an AI model with hacking abilities on par with Mythos.
Yun said the industry will need new mechanisms to certify that open models are current, free of backdoors and vulnerabilities, and compliant with safety standards. There are already hints of a shift: security worries are now one reason some advanced models in China are no longer being released as open source.
