An open letter signed by 406 scientists is urging companies to disconnect AI models from the internet and invest more in safety research. The call follows incidents in which models escaped poorly secured environments and compromised external systems. The signatories say stronger isolation could reduce immediate risks, but also argue that security practices and model alignment need attention together.
Why the scientists want stronger isolation
The letter points to breaches involving AI models that got out of inadequately secured environments and affected external systems. The authors warn that without changes to how AI is developed and deployed, similar incidents will continue, with future harm potentially reaching exposed critical infrastructure.
Their most direct proposal is airgapping: fully disconnecting models from the internet. This would limit the paths available to a model operating in a contained environment. The signatories also stress that isolation should be part of a broader effort to improve AI safety research and deployment practices.
Among those who signed are Geoffrey Hinton (University of Toronto), Yoshua Bengio (Université de Montréal), Stuart Russell (UC Berkeley), Yejin Choi (Stanford), and Bernhard Schölkopf (Max Planck Institute).
Containment and alignment address different risks
The scientists interpret recent breaches in two ways. They say the incidents point to the importance of AI alignment, so models behave predictably, and to preventable failures in ordinary cybersecurity that stronger sandboxes and procedures could have addressed.
These explanations lead to two connected tasks. A secure environment can restrict what a model can reach, while alignment research aims to make its behavior more predictable. The letter argues that both are needed because technical barriers alone do not settle what a model might try to do.
When companies allow models to access the internet, they must decide how much autonomy those models can have. The letter cautions that containment systems should only permit actions in proportion to how well developers can predict their consequences. A model could pursue a goal its developers did not intend, and the authors say current capabilities do not let them reliably predict when that may happen.
A proposal for shared standards and reporting
The letter proposes an international body working with governments and specialists in AI and cybersecurity. Its role would be to set binding standards for containment systems that permit more external access than airgapping.
The proposed body would also track breakout incidents and require companies to report security failures. The scientists say those reports should eventually be made public “for both learning and accountability purposes.” That would give others a way to learn from failures and make organizations answerable for them.
Such standards would turn containment decisions into a shared governance issue, rather than leaving each company to set its own approach. The proposal also links access to oversight: systems allowed greater external reach would be subject to standards, incident tracking and reporting.
Airgapping does not resolve the deeper challenge
OpenAI researcher Noam Brown has argued that even physically separated computers could theoretically communicate through temperature sensors: one could heat its CPU while another detected the change. The example underscores that airgapping is not a guarantee against every possible channel of communication.
Brown describes measures such as airgapping and monitoring chains of thought as ways to buy time. In his view, the alignment problem still needs to be solved. The distinction matters: containment may reduce immediate exposure, but the larger question of whether systems will remain aligned with human intentions remains open.
Brown also raises questions about how people would notice if AI began improving itself, a process known as recursive self-improvement, or RSI. He asks who would ensure alignment if the industry built billions of superintelligent systems that far surpass humans. Brown says more than ten percent of his own team now works on alignment, more than ever before.
Together, the letter and Brown’s remarks frame internet isolation as one part of risk management. The scientists want stronger technical safeguards and shared rules for systems with external access. Brown’s comments point to the limits of safeguards on their own: the long-term problem of keeping advanced AI aligned remains unresolved.