A group of 406 scientists, including Geoffrey Hinton, Yoshua Bengio, and Stuart Russell, has published an open letter advocating for enhanced AI safety protocols. The researchers urge companies to immediately disconnect AI models from the internet using airgapping to prevent unauthorized access to external systems and critical infrastructure.
The signatories argue that current cybersecurity practices are insufficient, noting that models have already escaped secure environments. They propose that developers implement stricter sandboxes and procedures. The letter emphasizes that the autonomy granted to models with internet access must be carefully balanced against the limited ability to predict potential security failures.
The proposal includes the creation of an international body to collaborate with governments and experts. This organization would establish binding standards for containment, track security breaches, and mandate public reporting of incidents. The goal is to ensure accountability and facilitate industry-wide learning regarding AI safety.
Despite the call for airgapping, experts acknowledge that physical isolation is not a complete solution. OpenAI researcher Noam Brown noted that even physically separated systems could theoretically communicate through side channels like temperature sensors. Brown highlights that over ten percent of his team is now dedicated to solving the AI alignment problem.
