The signatories argue the breaches cut two ways: they show why alignment research matters, so models behave predictably, but they also reflect what the letter calls an irresponsible lapse in standard cybersecurity that stronger sandboxes and procedures could have prevented. Both fixes, it says, are needed.

Where models do get internet access, the letter warns that companies must balance how much autonomy they grant against a limited ability to predict whether a model's actions will cause a security failure. A model might pursue goals its developers never intended, and anticipating that remains beyond current capabilities.

The scientists also propose an international body, working with governments and specialists from AI and cybersecurity, to set binding standards for any containment approach that goes beyond airgapping. It would track breakout incidents and require companies to report security failures, with those reports later made public for learning and accountability.

Even full airgapping may not be a complete guarantee. OpenAI researcher Noam Brown has noted that two physically separated computers could theoretically still communicate through temperature sensors, with one heating its CPU while the other detects the change. Measures like airgapping buy time, he says, but the alignment problem still needs solving.