US and China Must Cooperate on AI Safety, Says MIT Researcher at Beijing Summit

At a major artificial intelligence conference in Beijing's Zhongguancun district, a striking consensus emerged: the US-China AI rivalry risks overshadowing a more urgent shared threat.

The event, organized by the Beijing Academy of Artificial Intelligence, brought together computing pioneers including Whitfield Diffie (public-key cryptography co-inventor) and Turing Award winner Andrew Barto. But the central message was sobering—as agentic AI models grow more powerful, uncontrolled development could spawn cyberattacks and systemic failures that neither superpower can contain alone.

The Case for Collaboration#

Stephen Casper, an MIT computer scientist who participated via video, framed the dilemma starkly: "AI is a global technology with global benefits, global harms, and a consistent tendency for new capabilities to eventually proliferate." He cited research showing that international collaboration on AI safety outweighs national security costs—drawing a parallel to Cold War-era US-Soviet nuclear cooperation.

The cybersecurity landscape illustrates the point. Conference sessions highlighted emerging vulnerabilities in AI-generated code, novel attack vectors enabled by agentic tool use, and automated social engineering tactics that neither nation has mastered as a defensive counter.

Lin Yun, a professor at Shanghai Jiao Tong University, acknowledged that competition complicates alignment but stressed its necessity. "If different countries understand the risks in similar ways, it becomes easier to develop shared safety principles and technical standards," he told the author. "The key is to find areas where sharing can reduce systemic risk without exposing sensitive operational details."

The Open-Weight Model Dilemma#

The challenge intensifies as Chinese companies lead in releasing capable open-weight models—Moonshot's Kimi, Alibaba's Qwen, and Z.ai's GLM. The US has restarted its own push with Nvidia's Nemotron, but an inflection point looms. GLM 5.2, China's latest, reportedly includes frontier agentic and coding capabilities. And this week, Chinese cybersecurity firm 360 Security Technologies claimed to have built an AI model with hacking prowess matching Anthropic's Mythos.

This escalation stands against a backdrop of tightening US export controls. Washington has restricted chip sales and recently ordered Anthropic to block foreign nationals from accessing its most advanced models—a move the company answered by revoking universal access.