An AI application is an agent holding your credentials, reading text you didn’t write, and calling tools that cause side effects. That’s a different threat model from the web apps most of us grew up securing, and it’s the reason so many AI projects stall between demo and production. In this full-day guided walkthrough, Brian focuses entirely on how to build, review, and ship AI-powered applications and agents securely, using C#, GitHub, Microsoft Foundry, and Azure. This is Brian talking and demonstrating, not a hands-on lab. You’ll watch each pattern work, see what happens when it’s missing, and leave with reference code you can run on your own.

The day opens with the threat model, grounded in the OWASP guidance for LLM and agentic applications and the regulatory obligations that now apply, then follows the attack surface from the outside in. Brian starts with the supply chain: the dependencies AI tools suggest, the MCP servers and agent plugins your team installs, and how to verify what you build and ship. From there he turns to the coding agents themselves, showing how prompt injection reaches an agent through issues, pull requests, and comments, and how GitHub’s agentic workflows, managed agent permissions, and Azure DevOps policies contain it. The morning closes on hardening the application: guardrails and content safety in Foundry, gating tool calls, structured outputs, and treating evaluations and red teaming as security controls rather than quality checks.

The afternoon is about running securely. Brian covers identity for agents with Entra, keyless authentication, least privilege for tools, and MCP authorization done correctly. He then works through securing Foundry deployments end to end, from private networking and data boundaries to policy enforcement and API Management as the AI gateway, and compares that with self-hosted inference for workloads where residency or provenance demands it. The day closes with observability and incident response for probabilistic systems, responsible AI as an engineering discipline with a named owner, and a deployment matrix for Azure that shows which controls each option gives you for free. Throughout, Brian applies the disciplines from Essential DevOps, the book he co-authored with Ian Griffiths: small batches, review gates, and everything as code, now including your instructions to the agents.

This workshop assumes you’ve built with C#, used GitHub Copilot or a similar tool, and have at least a working understanding of models, RAG, and agents. It does not re-teach those foundations. The examples are .NET on Azure, but the threat model and the patterns apply to any language and platform.

You will learn:

• How to threat model AI applications and agents, and how to secure the supply chain behind them, from suggested dependencies to MCP servers and build attestations
• How to contain prompt injection and limit the blast radius of coding agents across GitHub and Azure DevOps
• How to give agents their own identity, harden and monitor AI workloads on Microsoft Foundry and Azure, and decide when self-hosting is the safer choice