A test that measures whether AI safeguards can prevent undesired outcomes when a model intentionally tries to bypass them.