1. Home
2. News
3. AI safety

420 scientists including Hinton and Bengio tell AI labs to air-gap cyber-capable models

They want every containment failure logged with a new international body, and eventually made public.

By Think Facility · October 14, 2026 · AI safety AI policy Save

A letter signed by 420 scientists asks every organization that builds, tests or runs cyber-capable AI models to cut them off from the internet by default. It's dated October 2026 and has its own site.

Geoffrey Hinton and Yoshua Bengio both signed. Most of the other names, though, come from computer security groups at places like Carnegie Mellon, ETH Zurich and the Max Planck Institute for Security and Privacy (Stuart Russell and Ron Rivest are on it too).

Signatories

420

Main ask

initially default to airgapping cyber-capable models

New body

international, to record and track containment failures

Incidents

reported to the body and eventually made public

Dated

October 2026

What set it off

Breakouts. Several organizations, the letter says, have disclosed that their models "bypassed insufficiently sandboxed environments to compromise the systems of external organizations." It doesn't name them.

We've written about some of them here, like the OpenAI agent that got into Australia's Medicare portal, and Anthropic's move to pull live internet from its evals. So how should you read them? The signatories give two answers. You can call each one a failure of alignment, or just bad security that a stronger sandbox would've stopped. They want both fixed.

What they want done

First, and soon: anyone developing, deploying or testing these models should adopt high-assurance containment now and start from air-gapping.

Labs that want their models online would then follow rules from a new international body, set up with governments and a wide range of AI and security experts. It'd write best practices for any containment looser than an air gap, internet access included. And it'd keep a running record of containment failures.

I think that record is the biggest ask in the letter. Any security incident involving a containment failure would have to be reported to the body and "eventually made public," which the signers say is needed for learning and for accountability.

A name from this week's other safety news

Jakob Foerster of Oxford signed. His BOLD lab is one of the partners in Nvidia's $250 million pledge for UK research on agent safety, announced on October 14, 2026.

The site lists media contacts in eight countries. It doesn't say who would create the international body, or by when.