Geoffrey Hinton among 415 scientists calling for joint effort for AI safety

Signatories called on companies to ‘exercise caution’ and ‘default to air-gapping cyber-capable models’.

More than 400 high-profile signatories, including Nobel laureate Geoffrey Hinton and Turing award winner Ron Rivest, are urging frontier model makers to work with governments to create an international standards body to improve AI containment and safety standards.

Experts, including Hinton, have backed previous calls warning against AI’s “unprecedented dangers”. In an open letter last September, experts raised caution around the escalating risks presented by advanced AI systems that had already exhibited deceptive and harmful behaviour.

This came months before OpenAI and Anthropic’s systems escaped containment during training and hacked real organisations. The two organisations were part of a group of 120 companies that signed an open letter this August calling for greater collective action on cyber defence. Leaders of the two companies have also made separate calls to pace frontier development.

AI technology has permeated across society in the few short years since bursting in popularity, becoming a key productivity and “creative” tool used by general consumers and enterprises alike. A fresh Eurobarometer survey from today (14 October) has found that 42pc of European SMEs and mid-cap companies already use AI, with another 21pc planning to do so in the future.

In their latest call, the 415 signatories called on companies to “exercise caution” and “default to air-gapping cyber-capable models” to physically isolate them from unsecured networks.

The scientists further want organisations to work with governments, and AI and cybersecurity experts to establish an international body that works for the “public good” to establish best practices for alignment and containment approaches. This body should also have a public record of containment failures, they demand.

“On the one hand, one can view these incidents as evidence for the importance of AI alignment to ensure models operate predictably,” the letter read.

“On the other hand, they can also be seen as an irresponsible lapse in standard cybersecurity practices, preventable with stronger sandboxes and procedures. We believe that both improved AI alignment, leading to more predictable model behaviour, and stronger sandbox technology are required.”

Without immediate changes in AI development and deployment practices, such incidents will continue to occur, and future incidents will be far more damaging to our exposed critical infrastructure, the experts warned.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.