Comment by nxpnsv

Is that not a recipe for adversarial training, thus ensuring increasing misalignment…?