GuardAgent introduces the first guardrail agent system designed specifically to protect LLM agents by dynamically verifying their actions against safety requirements.

• Creates task-specific safety plans from guard requests
• Converts safety plans into executable guardrail code
• Provides real-time protection beyond traditional text-focused guardrails
• Developed with EICU-AC benchmark for comprehensive evaluation

This innovation addresses critical security gaps in deploying autonomous LLM agents, enabling safer deployment in sensitive environments while maintaining functionality. GuardAgent represents a significant advance in building trustworthy AI systems that can follow safety protocols.

GuardAgent: Safeguard LLM Agents by a Guard Agent via Knowledge-Enabled Reasoning