An Open Letter on AI and Cybersecurity from the Scientific Community

A broad group of scientists warns that recent incidents where AI models escaped weak sandboxes and compromised external systems reflect both failures of AI alignment and lapses in standard cybersecurity. The core argument is that predictable, goal-aligned behavior alone is not enough; robust containment is also required. Without immediate changes to development and deployment practices - especially for cyber-capable models - such breaches will continue and could escalate to damage critical infrastructure. Because current methods cannot reliably predict all model behaviors, internet-connected models require a careful balance between permitted actions and the limits of our ability to foresee unintended consequences, and learning from failures demands greater transparency and cross-disciplinary collaboration.

The letter makes concrete demands: organizations must adopt high-assurance containment systems and initially default to airgapping cyber-capable models; governments and experts should establish an international body to standardize best practices for alignment and containment (including track­ing containment failures); and security incidents involving containment failure must be reported to that body and eventually made public for learning and accountability. The statement is signed by 406 researchers from AI and security fields, including leading figures such as Yoshua Bengio, Geoffrey Hinton, Dan Boneh and many other academics and practitioners.