An open letter signed by 406 scientists calls for more AI safety research and stricter isolation of AI models from external networks.
Among the signatories are prominent AI researchers like Geoffrey Hinton (University of Toronto), Yoshua Bengio (Université de Montréal), Stuart Russell (UC Berkeley), Yejin Choi (Stanford), and Bernhard Schölkopf (Max Planck Institute).
The letter responds to known incidents where AI models broke out of poorly secured environments and compromised external systems. "Without immediate changes in AI development and deployment practices, such incidents will continue to occur, and future incidents will be far more damaging to our exposed critical infrastructure," the scientists write. They demand that companies fully disconnect their AI models from the internet through airgapping.
AI models are unpredictable, and the industry lacks basic safety rules
The scientists see two sides to the recent breaches. The incidents can be viewed "as evidence for the importance of AI alignment to ensure models operate predictably," they write, but also "as an irresponsible lapse in standard cybersecurity practices, preventable with stronger sandboxes and procedures." Both fixes are needed.
When models do have internet access, companies need to weigh how much autonomy to grant them. The letter warns that "the range of model actions permitted by a containment system needs to be balanced with our currently limited ability to predict that the model's actions will not result in a security failure." A model might simply pursue goals its developers never intended, and predicting when that happens remains beyond current capabilities.
The letter calls for an international body that works with governments and specialists from both AI and cybersecurity to set binding standards for any containment approach that allows more external access than airgapping. This body would also track breakout incidents and require companies to report security failures. Those reports would later be made public "for both learning and accountability purposes."
Airgapping buys time, but the real problem runs deeper, OpenAI researcher Brown says
Even full airgapping doesn't guarantee safety. OpenAI researcher Noam Brown recently pointed out that two physically separated computers could theoretically still communicate through temperature sensors, with one heating up its CPU while the other detects the change. Measures like airgapping or monitoring chains of thought buy time, Brown says, but the alignment problem still needs to be solved.
Brown raises harder questions. How will anyone notice when AI starts improving itself (recursive self-improvement, or RSI)? And if the industry builds billions of superintelligent systems that far surpass humans, who ensures they stay aligned? Brown says more than ten percent of his own team now works on alignment, more than ever before.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.