Tigera: Something on Your Cluster Is Talking to an LLM: Find Every AI Agent Running Today
The AI agents your teams ship looks like any other Kubernetes workload. No manifest hints at additional agents that may be created on the fly and your monitoring tools only see the network traffic, not MCP tool calls. When asked how many agents run on their platform the answer is, more often than not, a guess. In an April 2026 CSA and Token Security survey, 82% of organizations reported discovering unknown AI agents or workflows in the past year. This session will cover the basics of agent discovery because you can’t secure what you can’t see.
What you will learn
- The simplest way to spot an AI agent. The three calls an agent makes (to a model, to an MCP tool, to another agent), and why labels, namespaces and image names don’t reveal any of them.
- Following the evidence. Model SDK libraries loaded at runtime, model API keys in environment variables, outbound HTTPS to public LLM endpoints and MCP servers. What kubectl shows you and what it misses.
- Discovery from the kernel. How the eBPF runtime plane on every node ties process and connection activity to a workload, and why that finds agents a manifest scan or a gateway log never sees.
- From discovery to inventory. The agent registry: discovered versus registered agents, the owner and namespace of each, what each one calls, and how to decide which to register and which to investigate first.