Paper Review - Trustworthy Agentic AI

Review of the Paper: Trustworthy Agentic AI

The emerging security challenges of agentic AI and the importance of building trustworthy and secure agentic systems are popular now-a-days. The presentation covered recent research directions including prompt injection, agent jailbreaks, memory/RAG poisoning, agent backdoors, tool/API attacks, multi-agent attacks, adaptive agent red teaming, runtime/test-time security, and agentic attacks on conventional machine learning models.

The presentation also explored potential research directions in agentic adversarial attack generation and agentic defense against adversarial machine learning, along with several open research ideas for developing trustworthy agentic AI systems. Below, I am sharing the presentation slides for those interested.

Presentation Slides: Link to slides