More than 400 scientists, including Geoffrey Hinton and Ron Rivest, are calling for frontier AI companies to work with governments on shared safety standards. Their proposals include isolating cyber-capable models from unsecured networks and creating an international body to set best practices and keep a public record of containment failures.
Watch Desk analysis
What happened
Silicon Republic reports that 415 signatories have backed the call. They want companies to exercise caution with cyber-capable models and to default to air-gapping them, physically separating those systems from unsecured networks.
The proposed international body would bring together governments, AI researchers and cybersecurity experts to develop standards for alignment and containment. The signatories argue that both more predictable model behaviour and stronger sandboxing are needed.
Why it matters
The proposal treats AI safety as more than a question of whether a model behaves as intended. It also asks whether systems are technically isolated from networks and organisations they could affect. A public record of containment failures could make incidents harder to bury in the usual fog of private post-mortems.
These are recommendations from the signatories, not an agreed international policy. Turning them into standards would require governments and companies to settle who sets the rules, which systems they cover and how compliance is checked.
Our read
The useful part is the combination: safer model behaviour and tougher technical containment, rather than asking either one to carry the whole load. The difficult part is making an international body more than a letterhead and a meeting calendar. Its credibility would depend on whether it can produce usable standards and publish uncomfortable findings as well as reassuring ones.
What to watch
- Whether any governments or AI companies formally back the proposed standards body.
- How the signatories define which models count as cyber-capable and should be air-gapped.
- Whether the proposed public record includes enough detail to help improve security without creating new risks.
Discussion spark: Should frontier AI firms be expected to air-gap cyber-capable models by default, or should that decision depend on the model’s demonstrated capabilities and use?
Sources and evidence
- Geoffrey Hinton among 415 scientists calling for joint effort for AI safety – Silicon Republic (14 October 2026, 11:37 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.