Agentic AI is already here, and many organizations have dozens of agents running. In Zentera's Agents of Change survey of 251 security leaders, 58% said they already operate more than 50. The question is no longer whether to govern them, but how. In this session, Nathanael Iversen, Zentera's Chief Evangelist and its representative to ONUG's agentic AI working groups, walks through what some of the largest enterprise adopters have documented through ONUG, how NIST and the Cloud Security Alliance are adapting frameworks auditors already know, and a practical architecture for keeping agents inside the work they were built to do.

Key highlights

  • Why Network Controls Aren't Enough: VLANs, subnets and firewall rules decide where an agent can connect, not what it does once it arrives. An agent that can reach a server can use every interface on it, so governance has to work at the level of intent, not just access.
  • What Enterprises Say Is Mandatory: In ONUG's poll of its large-enterprise members, 92% rated data guardrails mandatory before scaling agentic AI, alongside agent identity across trust domains and secure tool authorization.
  • Frameworks Your Auditors Already Recognize: NIST's proposed SP 800-53 control overlays for AI and CSA's MAESTRO framework, which ONUG has mapped to an agent's lifecycle, from creation and delegation through mutation and retirement.
  • Where Controls Have to Sit: On the endpoint where agents start, around the resources they reach, and in a layer 7 session controller that enforces each agent's declared intent. In early deployments, agents immediately spawned processes outside simple sandboxes.
  • Treat Agents Like Enterprise Software: The organizations running agents at the largest scale use approved lists, testing and change control, and confine projects in separate enclaves. It starts with an inventory: one manufacturer that expected 50 to 60 agents found 1,500 across 300 projects.

 


Resources mentioned:

"If you have a list of approved applications, you should have a list of approved agents. If you have a way to test applications, you should have a way to test agents."

— Nathanael Iversen, Chief Evangelist, Zentera Systems