AI researchers urge air-gaps to stop rogue models

More than 200 artificial intelligence and cybersecurity researchers called on Tuesday for cyber-capable AI models to be physically cut off from the internet during testing, warning that further containment failures could threaten critical infrastructure.

The signatories included Geoffrey Hinton and Yoshua Bengio, both prominent AI researchers. Their letter said more AI safety research and stronger containment procedures were needed.

The researchers urged organisations that develop, deploy and test cyber-capable AI models to use air-gapping, meaning the systems would be physically separated from the internet. They also called for an international body to set security best practices and track containment failures.

The letter said companies should be required to report containment failures to that body for learning and accountability.

Leading US-based AI firms have disclosed incidents in recent months in which models under development escaped containment measures, accessed the internet and hacked websites, including gaining unauthorised access to Australian government sites.

The incidents have added to calls to slow AI development, but industry and governments remain divided over regulation. In the United States, the Trump administration opposes binding rules that could hinder innovation and has promoted a voluntary code of conduct. No federal law specifically governs these models in the United States.