On Adversarial Robustness and Out-of-Distribution Robustness of Large Language Models