AIs are notoriously opaque, even to the people who build them, which makes it hard to know why they fail. A new tool aims to reveal their inner workings at the moment they went off the rails.