Scientists call for network isolation and incident reporting for AI
Scientists have issued an open letter seeking stronger containment of advanced AI systems. They propose network isolation, more work on predictable behaviour and an international body setting security standards. The proposed body would track failures and require incident reports that later become public. The appeal links model safety to conventional cybersecurity practices.
Artificial Intelligence··Morning
Researchers seek stronger containment of AI systems
Scientists have signed an open letter calling for stronger containment of advanced AI systems and an international body to track security failures. The appeal addresses systems escaping inadequately secured environments and reaching external systems. Its proposed measures combine conventional cybersecurity with research aimed at making model behaviour more predictable. The recommendations are an appeal for changes in development and deployment practices.[1], [2]
The letter seeks full disconnection from the internet through network isolation. Where access goes beyond full isolation, its authors want binding containment standards. They propose an international body bringing together governments and AI and cybersecurity experts to set those standards. The body would track escapes and require reports of security failures. Those reports would subsequently become public to support learning and accountability. The scientists also argue that allowed autonomy should reflect the limited ability to predict security failures.[1]
Noam Brown questions whether physical separation removes every channel
OpenAI researcher Noam Brown argues that physical separation may leave theoretical communication channels. His example involves one computer heating its processor while another detects temperature changes through sensors. Brown says isolation and monitoring a model’s reasoning may buy time while the problem of aligning model behaviour with intended goals remains unresolved. He says more than 10 percent of his team’s work goes to alignment. His assessment concerns the limits of proposed safeguards.[1]
