Securing AI Agents Without Building a New Security Stack

Your agents are making calls you didn't plan for. Run the same prompt twice and you can get two completely different call graphs: different LLM providers, different MCP tools, sometimes other agents entirely. Your security model assumed you knew the call graph ahead of time, and that assumption no longer holds. So when something goes wrong, it's not a 500. It's a leaked provider key, a write tool invoked by an agent that was only supposed to read, or a token bill nobody can put a name to.

Watching teams respond to this, we keep seeing the same move: build a whole new security stack for AI. That means JWT plumbing, custom headers, rate-limit middleware in every app, and one credential hack per provider.

You probably don't need to build any of it. The pieces already exist in the platform you're running today. They just weren't wired up for agent traffic. In this session, you'll learn:

• Agent identity isn't workload identity. There are three identities in every agent call, and collapsing them is where most platforms go wrong.
• We'll sort which agent security problems are actually new, and which ones you've already solved without realizing it.
• You'll see where enforcement has to live if you want "who made this call?" to have an answer, without touching agent code.

We'll demonstrate authorization, credential injection, tool-level access control, and telemetry tied to identity.