“Exposing AI model capabilities through a controlled API or user interface is an effective tool for understanding usage and enforcing safety boundaries.”