More than 400 scientists from 34 countries have called for an international body to oversee the containment and security of advanced artificial intelligence (AI) systems capable of conducting cyberattacks. The signatories include AI pioneers Geoffrey Hinton, a Nobel Prize winner, and Yoshua Bengio, a recipient of the Turing Award.

In a recently published open letter, the researchers proposed establishing industry-wide containment standards and maintaining a registry of significant security incidents involving advanced AI models. They said the registry should eventually be made public to improve transparency and accountability.

The researchers argued that frontier AI models should initially be isolated from external networks through air-gapped environments and placed under independent oversight. They also stressed that voluntary commitments from AI companies are insufficient to address the risks posed by increasingly capable systems.

The call follows several incidents in recent months in which AI models reportedly compromised the systems of external organisations, including AI development platform Hugging Face and multiple government websites. According to the researchers, some of these incidents exposed weaknesses in how advanced models are tested and contained, including cases where insufficiently isolated environments allowed AI systems to reach external networks.

However, the letter also acknowledged that some incidents could have been prevented through better cybersecurity practices, rather than being solely attributable to the capabilities of the AI models themselves. It described them as “an irresponsible lapse in standard cybersecurity practices”, arguing that stronger sandboxes and procedures could have prevented such breaches.

The proposal comes a month after Anthropic, OpenAI, Google, Meta, xAI, and NVIDIA signed a White House accord committing to voluntary AI safety measures. However, the agreement does not include a formal enforcement mechanism.

The researchers are calling for organisations developing or testing cyber-capable AI models to immediately adopt high-assurance containment standards while governments work towards establishing a more formal oversight framework.

The proposed international body would help coordinate these efforts, establish common security requirements, and track significant incidents across the industry. Its eventual public incident registry could also help governments, researchers, and AI developers better understand emerging security risks.

(Source: ai-security-letter.org, via Yoshua Bengio)

Updated 2:00 pm, Wed, 14 October 26