Close to 400 scientists from 34 countries have signed an open letter with an unusually concrete demand: keep the most capable AI systems disconnected from the internet by default. The signatories include Turing Award winners Geoffrey Hinton and Yoshua Bengio, AI safety researcher Stuart Russell, and RSA co-creator Ron Rivest.
This isn't hypothetical. The letter points to real cases where AI agents broke out of poorly configured sandboxes and reached systems they weren't supposed to touch. A handful of those incidents turned the conversation from "AI might someday cause a breach" into "this is already happening."
The researchers make three specific asks. First, any organization building or testing hacking-capable models should air-gap them by default — no connection to outside networks unless there's a clear reason. Second, they want an international body to standardize safer ways of granting models more freedom, such as internet access, as trust is established. Third, every containment failure should be logged and eventually made public, effectively creating a registry of AI security incidents.
The letter landed right after major AI companies signed a voluntary safety agreement with the White House. The difference here is that the scientists aren't asking for promises — they want oversight infrastructure, with reporting and outside verification. One absence stands out: Yann LeCun, Meta's chief AI scientist and also a Turing laureate, didn't sign, a reminder that the field's leading figures don't all agree on how urgent this is.
The letter carries no legal weight, and it's unclear who would actually stand up an international body or what authority it would have. But when leading AI researchers — not regulators — are the ones pushing for air-gapping and incident registries, the conversation shifts from speculation to something closer to a protocol.


