“All AI models, whether open or closed, should have custom guardrails applied by the user because the model provider's inherent judgments and values may not align with the user's.”