Agentic AI’s OODA Loop Problem

The essay analyzes how agentic AI systems executing OODA loops operate in inherently adversarial environments, making their observations, orientation, decisions, and actions vulnerable to manipulation (e.g., prompt injection, data poisoning, backdoors, reward hacking, and tool hijacking). It argues that current architectures lack semantic integrity and that speed and capability have been prioritized over verification, creating a security trilemma (fast, smart, secure) where integrity is sacrificed. The authors call for new integrity-centric architectures—beyond hallucination fixes and incremental training tweaks—to authenticate inputs, preserve trustworthy state, and verify tool use across the entire AI stack.