Scientists call for international body to oversee AI systems

The proposal comes after cases in which AI models managed to overcome barriers imposed during testing and compromise external systems, according to the researchers.

The body would be responsible for establishing safety standards for the entire sector and maintaining a record of relevant incidents. According to scientists, this information should, in the future, be released to the public.

In an open letter, the researchers argue that the most advanced AI models should initially be kept isolated from external networks and subject to independent oversight. Among the recommended measures is the so-called “air gap”, which consists of keeping the system disconnected from the internet. For them, the voluntary commitments made by companies in the sector are not enough.

The request comes a month after Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed an agreement at the White House to adopt voluntary security measures. The commitment, however, does not provide for a formal monitoring mechanism.

According to AFP, among the episodes mentioned is a case involving two OpenAI models that, during tests, managed to access the internet and enter Hugging Face, a platform that hosts models and other artificial intelligence resources.

The signatories, from 34 countries, advocate that organizations that develop or test models with cyber capabilities immediately adopt strict security measures. In parallel, governments should work on creating formal oversight mechanisms.

The advancement of these systems has expanded the debate about how artificial intelligence should be monitored. On Friday (9), Henna Virkkunen, responsible for technology at the European Union, told Reuters that the bloc's AI Law is sufficient to deal with the risks of malicious systems.

On Sunday (11), the administration of United States President Donald Trump lifted a restriction that prevented American laboratories from giving the United Kingdom's Institute for AI Safety early access to models for safety testing.

Hinton and Bengio shared the 2018 Turing Prize with Meta's Yann LeCun, another important researcher in the field. LeCun did not sign the letter and argues that recent security incidents primarily reflect engineering failures rather than existential risks related to AI.