Zero Trust reshaped network security by replacing implicit trust with explicit verification at every boundary. AI Ops needs the same shift. Most organizations run a portfolio of AI systems, LLMs serving chat, RAG pipelines on corporate knowledge, tool-using agents executing workflows, with controls that don't transfer cleanly and trust assumptions embedded everywhere: in prompts, in retrieved context, in tool calls, in agent handoffs.

This session applies Zero Trust principles across the AI Ops estate. We define a reference architecture for three patterns (P1 LLM, P2 RAG, P3 agent), map ZT tenets to each, never trust the prompt, verify the context, least privilege for tools, assume the model is breached, continuously verify, and present the minimum control set per layer.

The session covers a demo in which we walk through one pattern using both trust-permissive and Zero Trust (ZT)-aligned configurations. We close with a continuous verification model spanning model/prompt registries, eval pipelines, drift detection, and red-team automation, with SAIF, NIST AI RMF, MAESTRO, and OWASP Agentic Top 10 mapped to operational artifacts.