We develop a time-resolved PAC-Bayesian framework that connects adversarially robust generalization to learning rate, loss lanscape curvature, and stochstic gradient variance, providing a mechanistic explanation for robust overfitting in adversarial training.