Securing Agentic AI Systems: A Blueprint for Safe Autonomous Decision‑Making

Securing Agentic AI Systems: A Blueprint for Safe Autonomous Decision‑Making

Securing Agentic AI Systems: A Blueprint for Safe Autonomous Decision‑Making

As AI agents evolve from chatbots to autonomous actors capable of executing multi‑step workflows, new security challenges emerge. This session presents a practical architecture for securing agentic systems, covering identity propagation, tool‑level authorization, audit‑logging for agent actions, and isolation of long‑context memory. We explore strategies for preventing prompt injection, hallucinated action execution, and over‑permissioned tool use: failure modes increasingly observed in enterprise deployments. Drawing from multi‑region conversational AI systems handling millions of queries daily and contributions to NIST AI agent security guidelines, the session provides a prescriptive framework that blends traditional security engineering with modern LLM‑driven automation. Attendees will learn how to design agentic systems that are trustworthy, predictable, and compliant by default.