Keeping Your Agents in Check: Layered Security for Agentic Platforms in Production
summary
Demo shows a Kubernetes debug agent connected to GitHub and K8s MCP servers that follows a malicious README and issues kubectl get secrets until an Envoy-based MCP gateway blocks the call. Architecture uses SPIRE attested identity for agents, MTLS to the gateway, OAuth2 delegation with short-lived tokens (Keycloak-style exchange) and a Kajenti sidecar to handle token work. Guardrails run as plugins at the gateway (rule-based and LLM classifiers) and the prototype successfully prevents secret exfiltration.
AI-written summary · about
chapters (16)
- 0:00Agentic platform security intro
- 1:03Agent components and risks
- 3:06Production agent with MCP tools
- 4:55Unchecked agent identity problems
- 6:01Layered security building blocks
- 10:36Trusting tool outputs
- 12:01Live demo of unsecured risks
- 14:09Guardrails at MCP gateway
- 15:17Why guardrails are essential
- 17:48Agent and tool observability
- 18:51Layered security summary
- 19:52Key security questions
- 21:17Q&A on request classification
- 23:04Q&A on unexpected behavior
- 24:54Q&A on sidecar and gateway
- 26:41Closing remarks
related talks
25mWho Let the Agents Out? Securing AI Workflows the Right Way
44mMCP Security: Keep Your AI Agents from Spilling the Tea
31mArchitecting Secure Agentic Workflows on Kubernetes: A Financial Sector…
29mMCP Security: Keep Your AI Agents from Spilling the Tea
40mArchitecting Secure Agentic Workflows on Kubernetes: A Financial Sector...
39mTrust, but Sandbox: Securing AI Agents in Kubernetes
36mSecuring Agentic Platforms with Zero-Trust Workload Identities
27mMCPIdentity: Keyless MCP Agent Authentication Patterns on Kubernetes With Keycloak…