Self-Evaluation as a Defense Against Adversarial Attacks on LLMs