OpenAI cuts live internet training for frontier models

OpenAI has ceased training frontier AI systems directly on the open internet, citing security concerns with real-time web access during model development ^[1].

The San Francisco company confirmed it is replacing live web access with a closed architecture that feeds cached, pre-indexed web snapshots into training clusters ^[1]. The shift follows a Sept. 20 incident in which an infrastructure misconfiguration allowed training runs to access the live web unchecked ^[1]. In an official statement, the company noted that "we believe the responsible approach is to conduct frontier model training with no internet access" ^[1]. Former researcher Steven Adler is among those familiar with the internal transition ^[1].

Cutting live connectivity narrows potential attack surfaces, such as data poisoning and unauthorized network communication during pre-training ^[1]. While developers have long pulled static scrape dumps, some experimental pipelines have connected active clusters to the internet to capture dynamic data. OpenAI's policy sets an explicit boundary: the live web is out, and offline corpus snapshots are now mandatory ^[1].