Your Data Agent answers with the permissions of whoever asks. Your AI Foundry agent has its own identity in Entra ID. And the one connecting with an API key doesn't show up in any log. An agent's security doesn't live in a single place: it's spread across identity, network, workspace, data and model, and it's the architecture that decides which layer can fail without the agent seeing what it shouldn't.

In this session we build the defense-in-depth map for agents across Entra ID, Microsoft Fabric and AI Foundry: which control goes in each layer, what it protects you from and what it costs you. We'll see why a Contributor bypasses OneLake security (and so does their agent), when it pays to split workspaces and what you pay for doing it, how the controls change between an on-behalf-of agent and one with Entra Agent ID, and why the agent's instructions are not a control.

All with live demos: the leak, the fix with a consumption workspace, automated tests that plant decoy data and check, persona by persona, that the agent doesn't reveal it, and event correlation in Sentinel. You'll leave with the complete map to apply to your own architectures.