MIT CSAIL Principal Research Scientist Una-May O'Reilly discusses how she develops representatives that expose AI designs' security weaknesses before hackers do.