Hundreds of Leading Scientists Urge Immediate Internet Disconnection for AI Models

An open letter signed by 406 scientists is demanding immediate changes to how artificial intelligence systems are tested and deployed, urging developers to physically disconnect models from the internet through airgapping.

The petition features prominent figures in computer science and machine learning, including Turing Award recipients Geoffrey Hinton (University of Toronto) and Yoshua Bengio (Université de Montréal), alongside Stuart Russell (UC Berkeley), Yejin Choi (Stanford), and Bernhard Schölkopf (Max Planck Institute).

The researchers warn that current containment practices are failing, pointing to documented incidents where models breached poorly secured environments and reached external networks.

"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 signatories state.

Unpredictable Behaviors and Inadequate Sandboxes

According to the open letter, security lapses stem from two distinct challenges: shortcomings in standard cybersecurity and the unpredictable nature of advanced models. While stronger sandboxes and strict containment protocols can prevent basic security failures, the underlying behavior of AI systems remains difficult to anticipate.

The letter stresses that models can autonomously pursue goals unintended by their creators, an issue that existing alignment techniques cannot reliably foresee. Consequently, granting models open network access creates an unacceptable level of risk.

The researchers argue that the degree of autonomy permitted by any containment setup must be strictly weighed against the industry's limited ability to guarantee that a model will not trigger a security failure.

Calls for Global Oversight and Incident Reporting

To address the governance gap, the letter calls for the creation of an international oversight body working alongside governments, cybersecurity professionals, and AI researchers.

This body would be tasked with establishing binding security standards for any deployment system that permits network connections beyond complete airgapping. Additionally, it would track and document breakout attempts, requiring technology companies to report containment failures. These incident reports would eventually be made public to foster accountability and shared learning across the research community.

Containment Only Buys Time

While physical isolation provides an essential safeguard, experts acknowledge that airgapping is not a permanent solution to AI risks.

OpenAI researcher Noam Brown recently noted that even complete physical separation can leave subtle attack surfaces open. Brown pointed out that two isolated computers could theoretically communicate across air gaps via temperature fluctuations, where one system heats up its processor while the other detects the shift using thermal sensors.

Brown emphasized that containment measures and chain-of-thought monitoring primarily serve to buy time while the core alignment problem remains unsolved. He also highlighted looming questions surrounding recursive self-improvement (RSI)—how developers will recognize when systems begin upgrading themselves—and how to govern billions of potential superintelligent models. Brown noted that more than 10 percent of his own team is now dedicated to alignment research, marking the highest proportion to date.