Researchers propose international oversight of advanced AI containment

Geoffrey Hinton and Yoshua Bengio backed a letter calling for cyber-capable AI models to be isolated from external networks during testing.

The signatories want stronger safeguards against models escaping their testing environments. The Hindu reported more than 200 signatories, while Reuters's updated October 13 account put the number at around 400 researchers.

The letter proposes physically disconnecting models under test from the internet. This would limit access to external systems if software-based isolation proves insufficient.

It also calls for an international body to set shared containment standards and maintain a registry of significant failures. Developers and testing organisations would have to report breaches of their testing environments.

The researchers linked the appeal to company disclosures of models accessing external systems without authorisation. They warned of potential harm to critical infrastructure. That is the letter's risk assessment, rather than a finding that such damage has already occurred.

The proposed body has not been established. The appeal calls for independent oversight and transparency, while governments and industry remain divided over binding regulation.