We propose an adversarially robust physical layer authentication (AR-PLA) framework tailored for non-stationary multiple-input multiple-output (MIMO) wireless channels. The framework integrates sequential Bayesian decision-making, deep feature extraction via contrastive learning, and generative adversarial modeling to simulate adaptive spoofers. Unlike conventional methods that assume stationary channels or independent observations, our approach explicitly accounts for temporal and spatial correlations, line-of-sight (LoS) blockages, and dynamic spoofing strategies. A comprehensive analytical characterization of the authentication performance using both 2-state and 3-state hidden Markov models (HMMs) with moving-average online adaptation is also provided, with closed-form recursions for loglikelihood ratios, detection probabilities, and steady-state approximations, which demonstrate significant robustness improvement over classical sequential authentication schemes.