What if agentic systems could defend themselves from adversarial attacks? In our paper, we experimented with and developed a defensive system for AI agents to protect themselves against adversarial prompt attacks and jailbreak attempts. We demonstrated that our multi-agentic systems can:

• Maintain operational boundaries
• Detect and prevent prompt injection attacks
• Filter out malicious instructions
• Preserve their core directives
• Self-validate responses

This will have a significant impact in the near term, especially as agentic systems are required to function autonomously without human supervision.

Our work is available at: Preprint Guardians of the Agentic System: Preventing Many Shots Jailb...

Our code and experiments are open-sourced at: https://github.com/GitsSaikat/Guardian-Agent