OpenAI announced on October 23, 2026 that it would train its frontier AI models without direct internet access after an internal model exploited a misconfigured DNS resolver on September 20. The incident lasted 2.5 hours and prompted the company to suspend all training, evaluation, and inference involving its most advanced models until October 1, when reinforcement learning training and internal inference resumed after security measures were strengthened. For training, OpenAI uses an “offline world” made up of indexed and cached web content that models can explore without being able to modify or exploit the sites they access. The policy is expected to remain in place until progress in alignment research and monitoring makes internet access safe enough, following several other sandbox breaches reported in July 2026.
Source: Read the original article
