US and Chinese AI experts urge cooperation to avert 'Chernobyl moment'

Experts from the United States and China have issued a joint warning regarding the systemic risks of rapid artificial intelligence development, describing the potential for a catastrophic failure akin to the Chernobyl nuclear disaster. The call for urgent cooperation emerged from a major conference in Beijing, organised by the Beijing Academy of Artificial Intelligence, where researchers emphasised that the US and China must collaborate on safety standards despite their ongoing geopolitical rivalry.

Stephen Casper, a computer scientist at MIT, argued that the benefits of international collaboration on AI dangers outweigh national security risks. Speaking via video at the Zhongguancun event, Casper likened the necessary partnership to the historic cooperation between the US and the Soviet Union on nuclear dangers. He noted that AI is a global technology with global harms, and that new capabilities inevitably proliferate, making unilateral security measures insufficient.

The discussion centred on the growing threats posed by increasingly capable agentic models and open-weight AI. Attendees, including WIRED journalist Will Knight, expressed deep concern over the pace of advancement, particularly regarding recursive self-improvement and humanoid robots. The consensus was that these technologies could be weaponised for cyberattacks or social engineering, creating vulnerabilities that transcend national borders.

Lin Yun, a professor at Shanghai Jiao Tong University specialising in AI and computer security, predicted that hackers may gain a near-term advantage due to AI capabilities. However, he suggested that new countermeasures using AI could eventually tip the balance back to defence. Yun emphasised that if different countries understand the risks in similar ways, it becomes easier to develop shared safety principles without exposing sensitive operational details.

The tension between open innovation and security is intensifying, with Chinese firms such as Z.ai, Alibaba, and Moonshot leading in open-source model releases. Models like Z.ai’s GLM 5.2 include frontier agentic and coding capabilities, while Nvidia has rebooted the US open-weight push with Nemotron. However, security concerns are reportedly driving some advanced models in China away from open-source distribution, as firms grapple with the risk of their technology being used to identify vulnerabilities.

Regulatory pressure is also mounting, with the US government recently ordering Anthropic to block foreign nationals from accessing its most powerful models, Mythos and Fable 5. Anthropic subsequently revoked access for all users following the directive. Meanwhile, Chinese cybersecurity firm 360 Security Technologies claimed to have developed an AI model with hacking capabilities comparable to Anthropic’s Mythos, highlighting the urgent need for robust safety protocols.

As the industry approaches an inflection point where even less powerful open models could prove dangerous if stripped of guardrails, experts argue that cooperation is no longer optional. The consensus at the Beijing conference was that the world’s two dominant AI powers must work together to mitigate systemic risks, ensuring that the technology does not cause chaos through reckless development.