This OpenAI hub provides safety performance on four types of evaluations: harmful content, hallucinations, jailbreaks, and instruction hierarchy.