Around 400 scientists called for an international body to oversee AI containment standards. Geoffrey Hinton and Yoshua Bengio were among the signatories. The body would set industry-wide containment rules for advanced AI. It would also maintain a registry of major security incidents. The registry should eventually be made public. Frontier models should be air-gapped from external networks at first. Independent oversight of these models is proposed. Voluntary industry commitments are judged insufficient. The call follows recent incidents where AI models bypassed sandboxed environments. Organisations testing cyber-capable models should adopt high-assurance containment now. Governments are urged to work toward formal oversight. The EU AI Act is cited by some policymakers as sufficient for current risks.
What changed
Frontier AI models were not subject to uniform containment or incident reporting standards.
What it unlocks
Creating a global registry of significant AI security incidents.
- 400 scientists signed
- 34 countries represented
Sources