OpenAI implements offline frontier-model training
OpenAI disclosed that reinforcement-learning training resumed October 1 under hardened containment, using cached web content rather than live internet access.
ThereforeOpenAI says its containment tests led to software patches and tighter network restrictions.