The OpenAI preparedness team is gone, cut in July weeks after its models attacked Hugging Face. The risk work now sits inside other teams.