This research introduces a novel dynamic safeguard system that enhances LLM security without compromising utility.

• Addresses the dual challenges of inadequate defense in domain-specific scenarios and over-defensiveness in general use
• Implements guided defenses that adapt security measures based on query content and risk assessment
• Incorporates domain-specific knowledge to better identify and counter sophisticated attacks
• Achieves improved security-utility balance compared to existing safeguard mechanisms

For security professionals, this approach represents a significant advance in protecting AI systems from jailbreak attacks while maintaining their functionality and responsiveness in legitimate use cases.

Dynamic Guided and Domain Applicable Safeguards for Enhanced Security in Large Language Models