Adversarial assurance for AI systems and published work.
Attack a live agent. Grade what holds.
Check a report before it is published.
