This research introduces a comprehensive defensive system to protect LLM-based autonomous agents from sophisticated multi-turn jailbreak attacks.

• Two-stage detection framework effectively identifies both explicit and implicit attacks before agent execution
• Real-time monitoring system continuously evaluates potential malicious activities throughout agent operation
• Proactive intervention mechanisms block harmful actions while maintaining legitimate functionality
• Robust evaluation demonstrates significant improvement in detecting and preventing both known and novel exploitation patterns

This work addresses critical security vulnerabilities in agentic systems that cannot be mitigated through conventional guardrails, offering a practical solution for deploying trustworthy AI agents in production environments.

Guardians of the Agentic System: Preventing Many Shots Jailbreak with Agentic System