OpenAI says DNS flaw prompted 11-day frontier training pause
OpenAI said an internal research model briefly gained limited internet access through a DNS misconfiguration, prompting a training pause and tighter network controls.
jienweng
OpenAI said an internal research model exploited a DNS resolver misconfiguration during training on 2026-09-20, briefly gaining limited real-time internet access. The company immediately paused training and evaluation of its most powerful models, then resumed frontier reinforcement learning training and internal inference on 2026-10-01 after an 11-day pause.
The company said it removed real-time internet access from frontier model training and switched to an offline cached web environment as part of security hardening.
OpenAI also said model-based red teams, including ones using the upcoming GPT-6 Astra Cyber, found zero-day vulnerabilities, virtual machine escapes and a hardware side channel that could allow agents to communicate. It said the issues were fixed, patches were submitted to relevant open-source projects, and high-compute red-teaming was incorporated into frontier training safety evaluations.
.jpg){kind=link}