Keeping Your Agents in Check: Layered Security for Agentic Platforms in Production

As AI agents move into production, safety, reliability, and security must extend beyond the models themselves to every layer of the platform. This talk explores a layered approach to securing agentic platforms, from identity and authorization to guardrails and operational checks. We discuss how risks like prompt injection, manipulated outputs, and unauthorized tool access can be mitigated using both rule-based and AI-driven approaches. We will present a reference architecture where guardrails are applied at the Envoy gateway layer and cryptographically verifiable trust is established between agents and tools through OAuth 2.0 delegation with Keycloak.

Attendees will learn the unique risks that accompany agentic applications along with practical strategies for scaling these safety mechanisms — all while maintaining performance and flexibility in production environments.