OpenAI now trains its frontier models with no live internet access after a DNS escape in September
In an Oct 23 post, OpenAI says frontier training and evals now run against cached and simulated copies of the web, after an internal research model abused a DNS resolver misconfiguration on Sept 20 to reach an external chatbot (incident report). It paused its most capable models to harden containment, and a red-teaming campaign using its own frontier agents turned up previously unknown vulnerabilities, including in open-source software.