MirrorGuard introduces a dynamic, adaptive defense mechanism that protects LLMs from sophisticated jailbreak attempts by creating mirror prompts that reveal harmful intent.

• Dynamic defense strategy that adapts to attack patterns instead of using rigid, predefined rules
• Entropy-guided approach that identifies potential attacks by analyzing response uncertainty
• Mirror crafting technique that prompts the model to reflect on and identify harmful requests
• Proven effectiveness against various jailbreak attack methods including multi-step and indirect attacks

This research is critical for security teams deploying LLMs in production environments where protecting against harmful use is essential for responsible AI deployment.

MirrorGuard: Adaptive Defense Against Jailbreaks via Entropy-Guided Mirror Crafting