Large language models (LLMs) are revolutionizing how we interact with technology, but they're also vulnerable to a new type of security threat: prompt injection attacks. These attacks exploit the LLM's inherent trust in the input it receives, tricking it into disregarding user instructions and executing malicious commands hidden within the data. Imagine asking an AI assistant a simple question, and instead of a helpful answer, it directs you to a malicious website! That's the danger of prompt injection. Current defenses, like fine-tuning and prompt engineering, are either resource-intensive or ineffective. This research explores a novel defense strategy: turning the attacker's own tricks against them. By analyzing successful prompt injection techniques, researchers found that attacks and defenses share a common goal: manipulating the LLM to prioritize specific commands. The proposed defense method leverages this insight, using the structure of attack prompts as a “shield” to protect the LLM from malicious instructions. It then reintroduces the user's original query, ensuring the AI focuses on the intended task. Experiments on various open-source and commercial LLMs, including Llama 2, Qwen 2, and GPT models, show promising results. The new defense significantly reduces the success rate of attacks, sometimes even to near zero, while maintaining the LLM's accuracy and usefulness. This research demonstrates a clever, efficient approach to LLM security. While current work primarily focuses on text-based injection, future research aims to adapt these methods to defend against more complex, gradient-based attacks. As LLMs continue to integrate into our lives, developing robust security measures like this is vital to protect users and ensure trust in these powerful AI systems.