Governing the AI Fleet: How CISOs Can Stop Rogue Agent Actions Without Slowing Down Developers

Your engineers are already running AI coding agents. Those agents run shell commands, edit files, call internal APIs, and spawn sub-agents — most of it inside developer environments that never cross a network boundary, where gateways cannot see and sandboxes are too coarse to help. Most organizations cannot produce a list of the agents running in their environment, let alone what those agents are permitted to do. In this webinar session, Clint Hill, Product Technologist, shows and demos live a different approach: enforcement inside the agent harness itself, where every tool call is checked against policy before it executes rather than flagged after. We will cover how to start in observe mode with nothing blocked, how to turn on enforcement without breaking your engineering org, where this approach still has gaps, and how it runs in self-managed and air-gapped environments. Key takeaways - Why enforcement has to live inside the runtime loop. What gateways, sandboxes, directories, and SIEM each catch, what each structurally misses, and where in-harness enforcement fits alongside them rather than replacing them. - One policy set across every model and harnesses your team uses. Most organizations run more than one, and that is not reversing. Ory provides extensive coverage with 11 coding harnesses and 14 agent SDKs. - How to turn on enforcement without disrupting your engineers. Start in observe mode with every action evaluated and logged but nothing blocked, then write policy based on observed behavior via OpenTelemetry. - Prove what every agent did for auditing. Every action is recorded as allowed, denied, escalated, or approved, attributed to a named human owner, and exported to the SIEM you already run.