Over 400 scientists call for the most dangerous AI models to be isolated from the internet

Over 400 scientists, including artificial intelligence pioneers Geoffrey Hinton and Yoshua Bengio, have signed an open letter published on October 13, 2026. The text asks companies developing artificial intelligence systems to initially keep disconnected from the internet those models deemed capable of conducting cyberattacks autonomously.

The letter’s central proposal is the adoption of isolation from external networks — a technique also known as air-gapping — as an initial precaution for the most advanced models, before their eventual connection to the internet once risks have been assessed. This is a measure that, in computer security parlance, physically or logically separates a system from public networks to prevent uncontrolled remote use. The idea, the signatories explain, is to apply this caution to the most powerful models from the earliest stages of training and testing, so as to be able to observe their behavior before exposing them to a direct connection with networks open to the public.

The signature of figures such as Hinton and Bengio, both among the best-known names in artificial intelligence research and among the first to publicly raise the issue of risks linked to advanced systems, lends public weight to an appeal arriving at a time of growing institutional attention on the safety of the most capable models. In the same days, the European Commission launched an initiative to support the use of artificial intelligence in scientific research, while internal analyses at some companies in the sector continue to raise questions about the risks of a possible autonomous acceleration of model development.

The letter does not indicate a timeline or a mandatory enforcement mechanism: it is framed as a recommendation addressed to companies in the sector, not as a regulatory act or a legal obligation. It is therefore up to each company to decide whether and how to incorporate the proposal for isolation from external networks into its own security protocols, in the absence, at present, of any regulatory requirement to that effect.