Everybody is talking about governing AI agents. But before organizations can govern AI agents, they have to answer a more fundamental question: What non-human identities are already operating inside the environment?
Service accounts, workload identities, application identities, automation accounts, bots, API integrations, and increasingly AI agents are accessing enterprise systems and data every day. Unlike human users, however, these identities often don’t have the same governance lifecycle around them. They may not have an obvious owner, their privileges may not be regularly reviewed, their credentials may be years old, and in some cases nobody may remember exactly why the identity was created in the first place.
This is the Shadow NHI problem.
The rapid adoption of AI has brought new attention to non-human identities (NHIs), but NHIs aren’t new. Organizations have been accumulating them for years as applications, cloud services, integrations, automation platforms, scripts, and other technologies have required machine-to-machine access.
AI Agents Are Only Part of the NHI Problem
A service account created for an application five years ago may still be active today. An application registration may have powerful API permissions that were appropriate when it was deployed but have never been reviewed since. An automation account may continue operating even though the employee who created it has left the company. Credentials created for integrations may remain active long after their original purpose has disappeared.
AI adds another layer to this problem. Increasingly, an identity that appears to be another application, service account, or workload may actually support an AI agent or automated process capable of accessing enterprise applications and data. Not every NHI is an AI agent, but virtually every AI agent will need an identity and permissions somewhere.
That makes the first challenge surprisingly simple: Before you can govern these identities, you have to find them.
Discovering the Shadow NHI Environment
Identity platforms such as Microsoft Entra ID already contain a significant amount of information that can help organizations build an inventory of their non-human identities. Applications, service principals, managed identities, and workload identities can be examined along with information about ownership, credentials, permissions, creation dates, and recent activity.
But creating an inventory is only the beginning. Knowing that an organization has 300 service principals doesn’t tell the security team which ones matter. The identities need context. Organizations need to understand what each identity is, who owns it, what resources it can access, what privileges it possesses, when it was last used, when its access was last reviewed, and whether its credentials and behavior are consistent with its intended purpose.
This is also where AI-agent discovery becomes interesting. Organizations may not always have a convenient attribute that says, ‘This identity belongs to an AI agent.’ Identifying probable AI or automation identities may require looking at multiple signals, including naming patterns, permissions, ownership, creation information, and how the identity is being used.
YouAttest’s current NHI work, for example, already evaluates attributes such as credential state, owner presence, identity recency, and AI-agent naming heuristics as part of identifying and assessing non-human identities.
Not Every NHI Represents the Same Risk
Discovery quickly creates another problem: prioritization. Finding hundreds or thousands of NHIs isn’t particularly useful if the security team doesn’t know which ones deserve attention first.
Consider two service accounts. One has a documented owner, limited permissions, recently rotated credentials, regular activity, and evidence of a recent access review. The other has no owner, elevated privileges, an old credential, little recent activity, and no evidence that anyone has reviewed its access in years. Both are NHIs, but clearly they don’t represent the same level of risk.
This is why identity risk scoring can become an important part of NHI governance. Rather than treating every machine identity equally, organizations can evaluate signals such as ownership, credential age and expiration, privilege level, activity and dormancy, application permissions, publisher information, and other indicators to determine where investigation should begin.
The objective isn’t simply to label an identity as human or non-human. The more important question is: Which non-human identities represent the greatest risk to the organization?
Privilege Is Where Shadow NHIs Become Dangerous
An unknown automation account is concerning. An unknown automation account with administrative access is something entirely different.
Effective NHI governance therefore has to go beyond discovery and examine privilege. Organizations need visibility into directory roles, application and API permissions, cloud-resource access, consent grants, privileged-role assignments, and other paths through which a non-human identity can reach sensitive systems or data.
This is an especially important area as NHI security matures. Our current analysis, for example, identifies privilege exposure as one of the largest areas for expanded NHI risk analysis, including broader directory-role coverage, high-value Microsoft Graph permissions, Azure RBAC assignments, consent grants, and PIM status.
Ultimately, the security question isn’t whether the enterprise has service accounts – it certainly does. The question is which non-human identities have access they no longer need, shouldn’t have, or that nobody is actively governing.
The Ownership Problem
One of the simplest NHI risk indicators may also be one of the most important: Who owns this identity?
Human identities normally have organizational context. There is an employee, a manager, a department, a job function, and eventually a termination event. Those events provide natural opportunities for identity governance.
Non-human identities don’t necessarily have those boundaries. A developer creates an integration and later leaves the organization. A project ends but its service account remains. An application is replaced but its credentials are never removed. Over time, the organization accumulates identities that continue to have access even though accountability for them has disappeared.
This creates an orphaned NHI: an identity with access but without a clearly accountable human owner.
Ownership should therefore be a foundational part of NHI governance. Someone within the organization should periodically be able to attest that the identity is still required, that they are responsible for it, and that the privileges assigned to it remain appropriate. The current YouAttest NHI scoring work already treats the absence of an owner as a risk factor.
From NHI Discovery to NHI Governance
Organizations don’t necessarily need an entirely new security philosophy for AI agents and other NHIs. Many of the principles already exist within identity governance: establish ownership, apply least privilege, periodically review access, remove unnecessary privileges, and maintain evidence that these reviews occurred.
What changes is the population being governed.
A practical NHI governance lifecycle begins with discovery, followed by classification of the identity and assessment of its risk. Organizations can then establish accountable ownership, review and attest the identity’s access, remediate unnecessary privileges or obsolete accounts, and continue monitoring the identity as the environment changes.
Discover -> Classify -> Score -> Own -> Attest -> Remediate -> Monitor
This turns NHI security from a one-time inventory project into an ongoing identity-governance process.
AI Makes the Shadow NHI Problem More Urgent
AI agents dramatically increase the importance of solving this problem because agents may interact autonomously with applications, APIs, databases, cloud infrastructure, and sensitive enterprise data. As their capabilities increase, the identity and privileges assigned to an agent effectively determine what that agent can do.
But organizations shouldn’t wait until they have perfectly classified every AI agent before beginning governance. They can start with the identities they already have. Find the service accounts. Find the applications and service principals. Find the workload and automation identities. Determine who owns them, what they can access, and whether that access remains justified.
Then start looking for the identities that don’t fit the expected pattern.
Because before enterprises can solve AI governance, they first need to solve AI identity visibility. And there may already be far more non-human identities operating inside the enterprise than anyone realizes.
Join Us on #AuditTuesday
We’ll be digging into the Shadow NHI Problem on the #AuditTuesday GRC Podcast with Dave Ackley, Founder & CEO and Andrew Berkuta, COO and Zero Trust Expert at Goliath.
We’ll discuss how organizations can discover Shadow NHIs, determine ownership and privileges, identify the identities that present the greatest risk, and begin developing a governance strategy for service accounts, workloads, automation, and AI agents.
Bring your questions for the LIVE Q&A as we uncover what may already be hiding in your identity environment. Or contact us directly at @ YouAttest
- AI Agents Are Privileged Identities. Their Access Needs to Be Reviewed. - October 3, 2026
- AI Is the Unrealized Insider Threat: Why Enterprise AI Security Starts With Existing User Access - September 20, 2026
- The Service Account Wasn’t the Initial Breach. It Became Part of the Attack. - September 13, 2026
