Hundreds of AI researchers: cyberattack AIs must be trained on air-gapped systems

In an open letter, 487 AI researchers, including Geoffrey Hinton and Yoshua Bengio, said current AI training security is inadequate and called for strong sandboxing such as air-gapping for cyber-capable models, plus a new international body to set standards and report containment failures.

• AI worms and major incident fears drive safety calls at Dublin summit2026-10-14
• 406 scientists demand AI models be cut off from internet2026-10-13
• Microsoft President Says AI Less Regulated Than Milk2026-10-12
• Microsoft President Backs Deployment of AI Safety Evaluators2026-09-23
• AI Safety Won't Advance If We Rely On Two Companies, Says Smith2026-09-23