Wait, what was your old approach to model containment?

https://openai.com/index/securing-our-frontier-model-training-environment/