What are the most challenging aspects of deploying AI from security standpoint (internally for making sure AI doesn't go out of control)?