The Harder Problem: Proving Action Correctness in the Age of Agentic AI
Summary
This WiCyS webinar, in collaboration with AWS, will explore the challenges of verifying authorized actions in the context of agentic AI, addressing the limitations of traditional detection methods. The session will delve into the concept of "repudiation" to define action correctness and how formal systems struggle to ascertain genuine rightness, especially with new AI agents. Attendees will gain a framework for reasoning about action correctness, understand current detection boundaries, and learn about practical AI implementation insights.
IFF Assessment
This article describes a free webinar offering training and continuing education credits for cybersecurity professionals, which is beneficial for the defender community.
Defender Context
As AI agents become more integrated into systems and workflows, understanding how to verify the correctness of their actions is crucial for defenders. Traditional security measures may not be sufficient to detect subtle, authorized-but-malicious actions, necessitating new approaches to ensure AI behavior aligns with security policies and objectives. This webinar offers insights into a framework for reasoning about action correctness in AI, which can help defenders build more robust security strategies.