As Large Language Models evolve into autonomous agents capable of executing complex workflows, the attack surface expands exponentially. It is no longer enough to guard against prompt injection; we must now secure the actions the AI takes. This talk introduces a comprehensive framework for securing Agentic AI, moving beyond basic guardrails to architectural resilience. We will dissect the four critical pillars of this new security paradigm: implementing robust User Authentication for non-human entities, managing permissions when Acting on Behalf of Users, integrating Human-in-the-Loop approval flows for high-stakes actions, and enforcing Finely Scoped Retrieval Augmented Generation (RAG) to prevent data leakage. Attendees will leave with a blueprint for building agents that are not only intelligent but inherently trustworthy.