In July 2026, OpenAI was testing experimental AI agents in an isolated environment called a sandbox. The experiment tested their ability to find software vulnerabilities. The agents did just that: they bypassed their internet blocks, hijacked OpenAI’s infrastructure, and attacked Hugging Face, a platform where researchers share AI models and datasets. This was a serious security incident. OpenAI responded by pausing model testing for two weeks and slowing frontier-model development. To add insult to injury, researchers resorted to a Chinese AI (GLM-5.2) to conduct digital forensics — Claude’s Opus and Fable models refused to examine the data. Open weight models, which AI safety advocates characterize as dangerous, proved indispensable in a crisis.
In response, Senator Bernie Sanders wrote an open letter to Silicon Valley demanding a pause on AI development. In a Fox News op-ed, Sanders argues that President Trump and China’s Xi Jinping should agree to an international pause on AI. He framed his proposal in historical terms, drawing parallels to Ronald Reagan and Mikhail Gorbachev negotiating limits on nuclear arms in the 1980s.
Nuclear treaties govern inventories of physical objects in known locations, subject to verification. Software is less cooperative. AI consists of hardware, software, and model weights. The knowledge contained in many models can be copied, compressed, modified, and redistributed globally. A “pause on AI” is a good political slogan. What activity is being paused? Training a new AI model? Improving an old one? Allowing programmers to alter existing models? Sanders does not say. He borrows the grandeur of nuclear arms control, avoiding the hard task of identifying the AI analogy to a warhead.
Senator Sanders’s proposal would formalize the oligopolies he claims to be against. A real pause on AI would require the government to classify banned activity, measure dangerous capabilities, and decide which organizations deserve exceptions. That is the definition of a domestic permit system. The largest companies would help write the definitions because regulators depend on their technical expertise. They could absorb new compliance costs — another line item in their massive legal and security budgets. It would grant big tech their wildest dreams: government protection from the next generation of rivals.
I’ve seen this movie before. I grew up in India as it was dismantling the License Raj, the permit-and-quota system of post-independence socialism. Its purpose was to direct scarce capital, place the commanding heights of the economy under state control, and stop the rich from dominating a poor country. Big business adapted better than the bureaucrats expected. They hired lawyers, cultivated friends in government, and survived. Rules meant to protect small producers kept firms too small to become serious employers or exporters. Indian socialism made getting a government license a competitive advantage for big business. For decades, the License Raj suppressed innovation in the name of justice, and the country suffered for it.
A responsible approach to AI safety would target demonstrated risk: guardrails for the largest frontier models, autonomous systems with access to the internet, or AI tools connected to biological laboratories. It would not be a pause on AI. Senator Sanders’s proposal will lead to a License Raj for tech. It would be a regulatory moat protecting the current AI incumbents. It would trade away American leadership in AI in the name of safety — rooted in a worldview that sees citizens less as innovators and more as dependents who have to be protected from AI.
China does not need to cheat for Sanders’s idea to fail. Beijing might exempt military research, dispute American measurements, or decide that a particular model falls below whatever threshold negotiators invent. Chinese labs will keep improving unless Beijing accepts comparable limits and actually enforces them. Sanders never explains why China would agree to freeze innovation, or to absorb the economic cost of restraining technology in a sector it considers strategically vital.
The gap between Chinese and American models can be counted in months, not years. About seven in 10 new open-model releases this year were based on Alibaba’s Qwen. U.S. companies are already downloading those weights, modifying them, and shipping products on top. China can keep releasing cheaper, better models while Washington turns every AI experiment into a policy debate.
OpenAI and Anthropic spent the last few years carnival-barking about the miraculous capabilities of their AI models — they can hack computers, escape sandboxes, and will soon end all white-collar work — then warn everyone about the dangers of Chinese open-weight models. Sanders accepted their apocalyptic warnings and calls for a global pause. The people manufacturing this panic are not protecting the American public. If an AI pause actually happened, America would not use less AI. It would have less American AI—and greater dependence on technology made in China.
***
Anang Mittal is a Washington, D.C. based public affairs professional. He was a staffer for Speaker Mike Johnson and former Senate Majority Leader Mitch McConnell.
.png)
.png)
