L3 Networks, Inc.
Shadow AI and Agent Identity: Why Zero Trust Must Evolve

Blog

Shadow AI and Agent Identity: Why Zero Trust Must Evolve

AI agents can reach business data, call tools, and act on behalf of employees. Zero Trust has to extend to agent inventory, ownership, least-privilege access, and network boundaries.


Most organizations have spent years improving access controls for employees. Microsoft Ignite's sessions on Zero Trust for AI, agent discovery and agent sprawl highlighted why that work now needs to extend beyond people. AI agents can connect to applications, use tools and interact with information on behalf of the business. Those capabilities make agents useful, and they make unmanaged access a serious operational concern.

The Problem Starts With Visibility

An enterprise cannot enforce access policy against identities and tools it has never identified. New agents may arrive through sanctioned platforms, coding assistants, local tooling or experimental workflows. Some may have access to Microsoft 365 data or connected business systems. Others may operate through service identities that make ownership hard to determine.

The first implementation step is a documented inventory. For each agent, record the business purpose, the sponsor or owner, the platform where it runs, the systems it can access, the identity it uses and the conditions under which it should be retired. Inventory alone does not secure an agent, but it creates the accountability needed to manage one.

Every Agent Needs an Access Policy and a Network Boundary

An agent should receive the minimum permissions needed for an approved task. Access reviews should consider delegated permissions, application permissions, credentials, tool connections and the possibility of unintended data exposure. Lifecycle rules matter as much as initial setup: agents used for a short project should not retain standing access indefinitely.

Identity, however, is only part of the control model. The network paths available to an agent influence what it can reach and how quickly risky behavior can spread. Segmentation, outbound access rules, DNS and network visibility can help limit the blast radius without blocking legitimate workflows. An agent's access is only as safe as the environment it runs on.

A Practical Zero Trust Sequence for AI

For technology leaders, the roadmap is straightforward: discover agents, assign accountable owners, classify the data and tools they use, remove unnecessary permissions, apply identity and network controls, and review activity on a continuing basis. High-impact agents should be addressed before low-risk experiments. A 12- to 24-month improvement roadmap can connect immediate actions with platform and process changes.

L3's perspective is that agent governance belongs in the same conversation as Microsoft Entra, Conditional Access, networking and security operations. It should not become a separate pilot program that nobody operates after launch.


Take the Next Step: Parallax, L3's Microsoft Assessment Suite

A useful first step is to understand your existing Microsoft identity and app-permission posture. Parallax, L3's Microsoft Assessment Suite, offers a free read-only tenant snapshot, with deeper identity, Conditional Access and app-permission assessments available. A complete agent-discovery project would require separate scoping.

Start Your Free Parallax Snapshot

Related Resources

Let's talk

Take the Next Step: Parallax, L3's Microsoft Assessment Suite

Understand your existing Microsoft identity and app-permission posture with a free read-only tenant snapshot, with deeper identity, Conditional Access, and app-permission assessments available.