Alignment Proposal: Adversarially Robust Augmentation and Distillation — LessWrong