Beijing’s Zhongguancun district is a strange place to hear calls for détente. This is China’s answer to Silicon Valley, a warren of glass-and-steel towers where startups race to beat each other — and the world — to the next breakthrough. Yet at a major artificial intelligence conference here just over a week ago, the message from the stage was not about winning.
It was about survival. The conference drew computing legends.
Whitfield Diffie, the man who co-invented public-key cryptography, was there. So was Andrew Barto, a Turing Award winner for reinforcement learning. Sessions covered recursive self-improvement — the notion that AI models can rewrite their own code and keep advancing without human help — and humanoid robots.
But none of that was the headline. The headline was this: frontier AI’s cybersecurity and systemic risks have grown too large for the US and China to treat each other as enemies.
The researchers in Beijing were not arguing from altruism. They were arguing from cold, hard physics. A model trained in one country can be deployed in another.
A vulnerability in one system can be exploited everywhere. Borders mean nothing to code. Agentic models are the specific worry.
These are AI systems that can take actions in the digital world — not just generate text or images, but execute tasks, move files, send messages, make decisions. They are becoming more capable.
And if they are not secured, they could cause real damage. The conference was packed with sessions on this problem. Nobody was pretending it was hypothetical.
Stephen Casper, a computer scientist at MIT, spoke via video. His words were blunt.
“AI is a global technology with global benefits, global harms, and a consistent tendency for new capabilities to eventually proliferate,” he told me afterward. That proliferation is the heart of the problem. A security hole in a Chinese model could be used to attack American infrastructure.
A flaw in an American system could be turned against Chinese networks. The technology does not care about flags. Until now, the US has treated the AI race as a national security competition.
Export controls, sanctions, restrictions on Chinese researchers. The Beijing conference offered a different calculation.
The economic stakes are enormous, yes. But the cybersecurity stakes are existential. No one side can fix this alone.
The conference did not produce a treaty. No joint statement, no handshake deal.
What it produced was a clear message from the people who build these systems: the rivalry is dangerous. Not because one side might fall behind, but because the technology itself, left ungoverned, could turn on everyone. That is the takeaway from Zhongguancun.
Not a new algorithm. Not a faster chip. A warning.
Sources
- Beijing sets out plan to lead China's AI drive
- ChinaAMC Unveils Strategy for H2 as AI Shifts from Conviction …
