미국 AI 전문가가 본 중국 AI 현주소와 공통된 두려움

Comment Loader Save Story Save this story Comment Loader Save Story Save this story Just over a week ago, I attended a major artificial intelligence conference in Zhongguancun, Beijing’s bustling high-tech district. It was packed with fascinating sessions touching on everything from recursive self-improvement—the idea that models can tweak their own code and advance indefinitely—to humanoid robots. And it featured a few legends of computing, including Whitfield Diffie, co-inventor of public-key cryptography, and Andrew Barto, who won the Turing Award with Rich Sutton for his pioneering work on reinforcement learning. But I left with one takeaway above all else: The US and China should put their fierce AI rivalry to the side. Frontier AI’s cybersecurity and systemic risks are too serious to ignore, and increasingly capable agentic models could soon cause chaos unless the world’s AI superpowers can work together. “AI is a global technology with global benefits, global harms, and a consistent tendency for new capabilities to eventually proliferate,” Stephen Casper , a computer scientist at MIT who spoke at the conference via video, told me afterward. Until now, the US has largely viewed China’s AI advances as an economic and national security threat. Washington has imposed tight restrictions on chips and chipmaking equipment to stymie the country’s development of powerful AI. Most recently, the US government ordered Anthropic to prevent foreign nationals from accessing its most powerful models, Mythos and Fable 5, over national security concerns . In response, Anthropic revoked access for everyone. One company that was of particular concern, WIRED previously revealed, was a South Korean telecom giant with alleged ties to China . But the conference, organized by the Beijing Academy of Artificial Intelligence, reinforced the idea that both the US and China stand to lose if AI is developed too quickly and recklessly. As AI becomes more powerful, more agentic, and more intertwined with everyday life, the risks that it could be used to conduct cyberattacks or fail in catastrophic ways will only grow. Because the world’s two dominant AI powers are responsible for the most advanced models, cooperation between them feels like it will be crucial. Casper pointed to research showing that the benefits of international collaboration on AI dangers outweigh any national security risks that come from working together. He likened the current situation to how the US and the Soviet Union were forced to work together on nuclear dangers, even as they sought to out-stockpile one another. “One thing that almost everyone in AI can agree on right now is that AI doesn't need a Chernobyl moment,” Casper said. One day-long session highlighted the universality of the cyber challenges raised by more advanced AI. This includes new kinds of vulnerabilities in AI-generated code , novel ways of attacking systems enabled by agentic tool use, and automated methods for carrying out social engineering attacks. After another session, I spoke with Lin Yun, a professor at Shanghai Jia Tong University who does excellent work on AI and computer security. Yun told me he expects hackers to gain an advantage over the near term, but that new countermeasures, including novel uses of AI, should tip the balance 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 similar ways, it becomes easier to develop shared safety principles and technical standards,” he told me. “The key is to find areas where sharing can reduce systemic risk without exposing sensitive operational details.” Perhaps the most pressing question for both nations is how to balance openness with risk. Open-weight models have become crucial for research and innovation, with Chinese models proving popular in the US. But as these models advance, it will become more challenging to ensure they don’t help hackers identify security vulnerabilities and can’t be wielded as cyber weapons. Over the last few years, Chinese companies have taken the lead in offering highly capable open-weight AI models, like Moonshot’s Kimi, Alibaba’s Qwen, and Z.ai’s GLM. The US has rebooted its own open-weight AI push with models like Nvidia’s Nemotron . But we’re approaching an inflection point where even less powerful open models could prove dangerous if they’re stripped of guardrails. The latest model from China’s Z.ai, GLM 5.2, includes frontier agentic and coding capabilities, according to expert analysis . The next generation of open-weight AI models might be just as capable as Fable or Mythos. In fact, just this week, 360 Security Technologies, a top Chinese cybersecurity firm, said it had developed an AI model with hacking capabilities on par with Mythos. Yun said the industry will need to devise new ways of guaranteeing that open models are up-to-date, free of backdoors and vulnerabilities, and have met safety standards. One possible sign of things to come? A source at one of China’s leading AI companies, who asked to remain anonymous because they weren’t authorized to talk to the press, tells me that security concerns are one reason why some advanced models in China are no longer being released as open source. This is an edition of Will Knight’s AI Lab newsletter . Read previous newsletters here.