You've secured your canvas apps. Hardened your flows with Key Vault and service principals. Thought through authentication and authorization across your Power Platform tenant. But autonomous Copilot Studio agents introduce an entirely different attack surface.

Agents consume external knowledge, call actions on behalf of users, and execute through triggers that nobody manually invoked. They generate natural language responses that can expose information in ways a form field never could. And prompt injection, where malicious instructions are embedded in content an agent reads, is a very real attack vector that many Power Platform practitioners have never had to account for.

Building an autonomous agent is the easy part. Building one that can be trusted is considerably harder. Securing agent workloads requires a different way of thinking. Vivian and Yannick will walk through the threats unique to AI-driven workloads and the security decisions that help mitigate them. Expect practical guidance, real-world examples, and a structured approach to evaluating agent security before go-live.