Zero Trust Agent Systems that Pass Audits and Still Ship

Most agentic AI demos assume a greenfield environment. In a real enterprise, agents run inside strict boundaries where security, compliance, and incident response are non-negotiable. The core problem isn’t just hallucinations, it’s that tool-using agents can quietly become a new insider with the ability to query data, call APIs, and make changes faster than traditional controls were designed for.

This talk shares a practical blueprint for building agent systems with a Zero Trust mindset. I’ll walk through an architecture that treats agents as first-class identities, enforces least privilege at the tool layer (not the prompt layer), keeps context and retrieval bounded, and produces audit-ready observability so teams can explain what happened and prove control effectiveness. The session is grounded in patterns I use in production environments where SOC 2, ISO, and similar controls matter.