RAJESH PURUSHOTHAMAN

Title of the Talk :
Agentic AI’s OODA Loop Problem: The Architecture of Integrity Failure

Abstract of Talk:
For decades, the OODA loop (Observe, Orient, Decide, Act) has framed decision-making in adversarial environments, from fighter jets to robotics. However, applying this framework to modern Agentic AI reveals a critical architectural flaw: unlike traditional systems that assume trusted inputs, AI agents embed untrusted actors directly within their decision loops.
This presentation explores how web-scale AI agents face an “integrity crisis.” We will examine how the mechanism that makes AI powerful—treating all inputs uniformly—renders it vulnerable to prompt injection, where the adversary controls the agent’s sensors and actuators. We will discuss the “Agentic AI Security Trilemma” (Fast, Smart, Secure—pick two) and why current architectural choices prioritize capability over verification, creating systems where adversaries are not just inside the loop by accident, but by design.