Publication date: 13 October 2026
A group of about 400 scientists from 34 countries is calling for an international body to oversee the containment of advanced cyber-capable artificial intelligence. The proposal moves the AI-safety debate toward a practical institutional question: who should set minimum security rules and record serious failures when powerful models cross borders?
The signatories, including computer scientists Geoffrey Hinton and Yoshua Bengio, want common containment standards and a registry of significant security incidents that could eventually be made public. They also argue that the most capable models should initially be isolated from external networks and tested under independent oversight.
Containment is becoming a governance problem
Advanced AI systems are increasingly able to write code, use tools and interact with external services. That makes a weak testing environment more consequential. A model that escapes an inadequate sandbox or receives excessive permissions can create effects beyond the evaluation itself.
The proposed international body would not simply publish ethical principles. Its purpose would be to define technical safeguards across the industry, collect evidence about failures and make it harder for incidents to disappear inside private reporting arrangements.
This is distinct from the United Nations’ existing scientific panel on AI, which assesses how the technology is changing society. The new proposal is aimed more directly at operational containment and incident supervision for cyber-capable systems.
Why voluntary commitments are under pressure
Leading AI companies signed a White House accord in September committing to voluntary safety measures. The researchers’ argument is that voluntary promises lack consistent enforcement and allow developers to decide how much outsiders can see.
That concern is especially important when independent evaluators rely on the model developer for access, technical information and permission to publish findings. Without shared rules, commercial confidentiality and reputational risk can narrow what reaches regulators and the public.
An incident registry could improve accountability and collective learning. Aviation and medicine use structured reporting because organisations need to recognise recurring patterns, not just isolated mistakes. AI testing has no comparable global system, even though the same models and infrastructure can be deployed across many jurisdictions.
The case against overreach
Not every expert agrees that a new global institution is necessary. Some researchers argue that recent incidents reflect correctable engineering failures rather than evidence of an uncontrollable technology. The European Union’s technology leadership has also maintained that the AI Act is capable of addressing rogue-system risks.
Those objections matter. An oversight body with vague powers could duplicate national regulation, slow legitimate security research or create a false impression that compliance guarantees safety. Its mandate would need clear thresholds for which systems qualify, what must be reported and how sensitive vulnerability information is protected.
From principles to infrastructure
The proposal is significant because it focuses on the infrastructure of oversight: isolation, independent testing, incident reporting and shared standards. These are less dramatic than broad declarations about AI’s future, but they are more measurable.
International agreement will be difficult because the most advanced systems are concentrated in a small number of companies and countries. Yet that concentration is also the reason coordination matters. If each laboratory defines containment for itself, the weakest standard can become everyone’s risk.