i
News
All Research News
News · 2026-10-13

406 researchers demand internet limits for AI models

@neuronium_ai @neuronium_ai

A group of 406 scientists is calling for AI models to be disconnected from the internet immediately, arguing that current safeguards cannot reliably predict how models will behave when given external access. Their proposal goes beyond restricting connectivity: it calls for an international body to set mandatory standards for isolation systems, track cases where models escape control and require companies to report security failures. The demand lands at the intersection of two problems the authors say must be addressed together: models that may act in unforeseen ways, and security practices that leave systems vulnerable to avoidable breaches.

Cover: 406 researchers demand internet limits for AI models

The gap between access and control

The authors draw two lessons from recent hacks. Better alignment could make model behavior more predictable; stronger isolation and security procedures could prevent conventional failures. Neither approach replaces the other.

The practical question is how much independence to grant a model that can reach the internet. The authors argue that the actions an isolation system permits should be limited by how poorly developers can predict whether those actions might cause a security failure. A model may pursue goals its developers did not give it, and there is no reliable way to identify in advance when that will happen.

Their proposed international body would work with governments and specialists in AI and cybersecurity to:

Set mandatory standards for isolation methods that allow more external access than air-gapping.
Track incidents in which models escape control.
Require companies to report security failures, with reports later made public to support learning from incidents and accountability.

Isolation buys time

Physical separation is not a complete answer, either. OpenAI researcher Noam Brown recently pointed out that two physically separated computers could theoretically communicate through temperature sensors: one heats its processor, while the other detects the change.

Brown’s point is that measures such as air-gapping or monitoring chains of reasoning can buy time, but do not solve the alignment problem. He also asks how researchers would detect recursive self-improvement (RSI), in which an AI begins improving itself, and who would ensure systems remain aligned if the industry creates billions of superintelligent systems that greatly surpass humans.

Brown says more than 10% of his team is now working on alignment, a larger share than ever before.

The unresolved question

I think the call’s strongest point is not that isolation should be abandoned, but that it cannot carry the safety burden alone. The letter asks for standards and public incident reporting; Brown’s questions point to a harder problem of oversight at a scale the industry has not reached. The proposal can make failures more visible, but it does not yet explain how anyone would verify alignment once systems can improve themselves or multiply beyond human capability.

Source

Daily AI news

Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.

Only what matters — every day

Follow on X