AI agent security roles: who owns each layer
Each layer of AI agent security has a natural owner: endpoint and browser sit with IT or endpoint security, identity with the identity team, the gateway with platform or network engineering, and the cloud workspace with the business platform admins. Security sets policy across all five and reviews results.
Facts read on .
Who owns the endpoint layer, identity layer and gateway layer?
Endpoint security or IT owns the endpoint layer, the identity team owns the identity layer, and platform or network engineering owns the gateway layer. Security owns the policy that all three enforce.
The owner is whoever already runs the install path. Vendors publish MDM and EDR paths for endpoint tools, so the team that runs the MDM runs the rollout.
Write the owner next to each layer in the layer map. If two teams each think the other owns the browser extension, nobody patches it. A named owner and a named backup close that gap.
Sources: Pillar Security: Agentic Endpoint; Operant AI: Endpoint Protector.
Who owns the browser and cloud workspace layers?
Endpoint or IT teams own the browser layer, as it ships as an extension or managed browser. Platform admins own the cloud workspace layer.
For Copilot Studio, Microsoft documents that admins use data policies in the Power Platform admin center to govern agent capabilities, so those admins hold the controls.
Business-unit makers build many cloud workspace agents. They need training and clear limits, since they are not security staff. Admins can use data policies to limit connectors and channels before a maker builds.
How do security, IT and engineering share the work?
Security writes the policy and reviews, IT and identity teams deploy, and engineering owns the agents it builds. A shared register names one owner per agent and per control.
NIST's AI RMF puts governance first, with named roles and accountability, which this split follows.
Hold a short monthly meeting between the layer owners and the policy owner. The agenda is the monthly report, the open gaps and the decisions needed. Keep it to thirty minutes.
Sources: NIST AI Risk Management Framework.
Who owns each agent?
Every agent needs a named human owner, because incident response starts there. Check Point says its discovery records each agent's owner, model, tools and MCP servers.
Microsoft's Agent 365 registry lists agents with owner, usage and risk. Orphaned agents should be disabled.
When an owner leaves, reassign their agents before the account is disabled. An agent that loses its owner keeps running with its credentials, which is how orphaned agents appear.
Sources: Check Point: AI agent security; Microsoft Agent 365 overview.
What should the policy owner decide?
The policy owner decides which agents and tools are approved, which actions need a human, and who may grant exceptions. Keep these few and written.
Conditional Access in Entra shows the pattern: policies are if-then rules that combine signals and are owned by one team.
Record who may approve each class of agent and tool, and keep the approvals in a log that auditors can read. Approvals tied to named roles survive staff turnover.
How do you avoid gaps between owners?
Gaps appear at the edges: a browser extension that reaches a SaaS agent, or a token issued by one team and used by another. Review those edges quarterly.
Use one inventory across the layers, and make each control owner sign off on the layer map.
Include the vendor relationship in the ownership table: who holds the contract, who gets the security notices and who reviews the documentation each quarter.
Sources: Zero Trust overview, Microsoft Learn.
Sources
- Pillar Security: Agentic Endpoint (pillar.security), read
- Operant AI: Endpoint Protector (operant.ai), read
- Microsoft Copilot Studio security and governance (learn.microsoft.com), read
- NIST AI Risk Management Framework (nist.gov), read
- Check Point: AI agent security (checkpoint.com), read
- Microsoft Agent 365 overview (learn.microsoft.com), read
- Microsoft Entra Conditional Access overview (learn.microsoft.com), read
- Zero Trust overview, Microsoft Learn (learn.microsoft.com), read