Large language models (LLMs) have shown remarkable capabilities, but they can also generate unsafe or harmful content. Ensuring LLM safety is a critical challenge, but new research suggests a surprising solution: self-evaluation. Researchers have found that pre-trained LLMs can effectively identify unsafe inputs and outputs, acting as their own security guards. This self-evaluation defense doesn't require costly fine-tuning or proprietary APIs. The LLM acts as both the generator and the evaluator, determining if content is safe or unsafe. The results are promising, showing a significant reduction in the attack success rate of adversarial prompts designed to elicit harmful content. In tests, the self-evaluation approach outperformed existing commercial content moderation APIs and other defense mechanisms. One intriguing finding is that while adversarial attacks can sometimes bypass an LLM’s safety mechanisms during content generation, the same model can still identify the unsafe input or output in a separate evaluation step. This suggests that adversarial attacks don't entirely break the model's understanding of safety. Researchers also explored ways to attack this self-evaluation defense. While they discovered methods to fool both the generator and evaluator simultaneously, the overall defense still proved stronger than using the generator alone. While more research is needed to explore the full potential and limitations, self-evaluation offers a simple, efficient, and robust approach to enhance LLM safety. This innovative technique could prove vital as LLMs become more integrated into everyday life.