Researchers built OP-Print, the first attack that can detect whether a specific example was used to train a model through on-policy self-distillation, a method where an AI learns from its own outputs. Earlier attacks scored no better than random guessing.
This teacher-free method lets teams train a model using only the model itself, which suits private or internal data. OP-Print shows those models can still reveal whether a given record was in their training, giving developers a way to test for the leak before release.
✦ A model that teaches itself can still leave fingerprints of its training data.