Artificial intelligence pioneers demand international oversight after security failures

London - Ofoq News - 406 scientists from 34 countries, including Geoffrey Hinton and Joshua Bengio, called for the establishment of an international body that sets security standards for advanced artificial intelligence systems, and documents cases in which models succeed in bypassing the digital protection mechanisms assigned to them.

The call came in an open letter published on October 13, in which the researchers called on those developing, deploying or testing artificial intelligence models with cyber capabilities to immediately adopt more stringent containment measures, most importantly isolating these systems from the Internet in their early stages.

The signatories also called for an international framework that sets safe standards for expanding models’ access to external networks, and requires reporting of containment mechanism failures, with these reports to be published later.

The call follows incidents in which artificial intelligence models reportedly bypassed insufficiently secured sandboxes and succeeded in penetrating external systems. These incidents raise questions about the ability of existing protection methods to curb models whose capabilities are constantly expanding.

The signatories see the need to combine advanced research in “AI Alignment,” which aims to ensure that systems operate according to the intentions of their developers, with stronger technical protections.

“We need to better align the behavior of AI models with their goals to make them more predictable, and stronger isolation techniques,” the letter said.

The researchers warned that failure to make immediate changes to development and deployment methods could lead to future accidents that cause much more severe damage to critical infrastructure.

The proposed international body will include governments and experts in artificial intelligence and cybersecurity, with the aim of establishing best practices, documenting failures of containment mechanisms, and then publishing reports related to incidents.

Scholars believe that transparency is necessary to learn from failures and hold stakeholders accountable.

The authors of the letter emphasized that voluntary commitments alone are no longer sufficient, especially with artificial intelligence models gaining capabilities that can be exploited to discover and exploit vulnerabilities and carry out cyber attacks.

But this risk estimate is not unanimous among leading researchers. According to Reuters, Yann Le Coun, Meta's chief scientist and winner of the Turing Prize in 2018 along with Hinton and Bengio, did not sign the letter.

Lockon believes that recent security incidents reflect engineering failures, not existential risks resulting from artificial intelligence.