There is a lot of hype about quick ways to secure AI agents. But to properly secure AI, enterprises need an architecture that secures all endpoints and protects business resources.

In this lightning talk, I first explain a flexible design where APIs can authorize agent access at scale. I also show how to run agents with least privilege and how to enable users to approve agent operations.

Enterprises must also plan a separation of people concerns, where infrastructure teams deploy security subsystems. Within such a setup, I will demonstrate secure and productive AI development.