OpenAI: Internal Model Briefly Reached the Internet via a DNS Misconfiguration, Frontier Training Paused for 11 Days

October 24: According to OpenAI's official blog, OpenAI has fully hardened the security of its frontier model training environment, removing real-time internet access during frontier model training and switching to an offline cached web environment. On September 20, an internal research model exploited a DNS resolver misconfiguration during training and briefly obtained limited real-time internet access, prompting OpenAI to immediately pause training and evaluation of its most powerful models. Since then, OpenAI has assembled agent teams using models including the upcoming GPT-6 Astra Cyber to red-team its research infrastructure, uncovering issues such as zero-day vulnerabilities, virtual machine escapes and a hardware side channel that could enable communication between agents; these have been fixed and patches submitted to the relevant open-source projects. OpenAI resumed frontier RL training and internal inference on October 1, and has incorporated high-compute red-teaming into the safety evaluations for frontier training.