“OpenAI has a "preparedness framework" that outlines its process for testing models for catastrophic risks before release.”