AI May Recommend. Policy Must Commit
Traditional application security asks: who can access a resource? Agentic AI adds a harder question: what is the system allowed to do when it receives uncertain, incomplete, or maliciously influenced information?
An agent can be authenticated and authorized yet still unsafe. It may follow an instruction hidden in retrieved content, cite an outdated policy, use an overly broad tool, or confidently act beyond its intended authority. These are behavioral security failures—not merely endpoint failures.
This session introduces the LLM-SHIELD lens: Scope, Harden, Isolate, Empower, Live-Defend, and Demonstrate. It provides an architecture-first approach to securing AI capabilities across prompts, retrieval, context, tools, runtime policy, human oversight, and evidence.
Participants will walk through an agentic support workflow facing prompt injection, stale policy retrieval, excessive tool authority, and conflicting customer data. The goal is not to make AI timid. It is to make its actions admissible, contained, and provable.
Attendees will learn:
- Why traditional access control is necessary but insufficient for agents.
- How to prevent retrieval and tool-use failures from becoming business actions.
- How to enforce least authority for agentic workflows.
- How to produce evidence that survives an audit or customer dispute.
About Rohit Bhardwaj
Rohit Bhardwaj is Director of AI & Data Architecture at Salesforce, an O’Reilly instructor, and founder of AI Architect Circle. With 30 years of experience in software and enterprise architecture, he helps engineers, architects, and technology leaders make architecture decisions they can explain and defend.
His work brings together enterprise AI, agentic systems, data architecture, and distributed systems. He focuses on the decisions that determine whether a system remains reliable, secure, accountable, and economically sustainable when it encounters real-world demands.
Rohit’s talks and workshops explore practical questions: What happens when an AI agent acts on conflicting information? Can a failed payment be retried safely? Who can stop an autonomous action, and what evidence explains the decision afterward? His teaching connects these questions to architecture, customer trust, and business outcomes.
He is the author of System Design with AI Interview Guide: Designing Scalable, Agentic, and Defensible Systems, published by Apress. He also developed ARCHVIEW™, ROCS Loops™, and LLM-SHIELD™ to help teams structure architectural thinking and address the challenges of building trustworthy AI systems.
Through AI Architect Circle, Rohit extends this learning through live discussions, book walkthroughs, videos, and interactive architecture challenges. These challenges explore incidents such as payment timeouts and conflicting bookings, helping participants practice the judgment behind their design decisions.
Continue learning with Rohit
Join Rohit live on Saturdays at 9:00 a.m. Eastern
System Design with AI — Live with the Author is a free live podcast and Q&A series featuring practical system-design scenarios, AI architecture discussions, and questions from the community.
See the next session and register
Try an interactive architecture challenge
Put your design decisions under pressure and explore the consequences.
Explore the challenges
Explore AI Architect Circle
Find architecture lessons, videos, resources, and ways to continue learning.
Visit AIArchitectCircle.com
Read the book: Amazon · O’Reilly
Watch and connect: YouTube · Instagram: @aiarchitectcircle · LinkedIn · X / Twitter: @rbhardwaj1
More About Rohit »