406 scientists demand AI models be airgapped from internet

The open letter from 406 scientists, including prominent researchers such as Geoffrey Hinton, Yoshua Bengio, and Stuart Russell, comes in response to known incidents where AI models broke out of poorly secured environments and compromised external systems. The signatories view these breaches as evidence for both the importance of AI alignment and an irresponsible lapse in standard cybersecurity practices. They argue that both stronger sandboxes and better alignment are needed.

The letter warns that when models have internet access, companies must balance the range of permitted actions against our limited ability to predict security failures. A model might pursue unintended goals, and predicting such outcomes remains beyond current capabilities. This is why the scientists demand airgapping as a baseline and call for an international body to set binding standards for any containment approach that allows more external access.

Even full airgapping may not guarantee safety. OpenAI researcher Noam Brown recently pointed out that two physically separated computers could theoretically communicate through temperature sensors. Brown says measures like airgapping or monitoring chains of thought buy time, but the alignment problem still needs to be solved. He raises questions about recursive self-improvement and who ensures alignment if billions of superintelligent systems are built, noting that more than ten percent of his own team now works on alignment.