"Select all squares with traffic lights, bicycles, or crosswalks."
Sigh.
In 2026, completing a CAPTCHA feels a bit like flashing your human ID at a club that's 97.8% robots. That isn't a random statistic: the Cloud Security Alliance puts the average enterprise's machine-to-human identity ratio at 45:1—a stark increase from 17:1 in 2023.
The irony is worth sitting with. Security teams already spend an enormous amount of effort making sure it's a human at the door (and the right human at that), while the non-human identities (NHI) inside—the service accounts, bots, and AI agents the organization deployed on purpose—often go unclassified, unowned, and unmonitored.
This article leverages guidance from the Open Worldwide Application Security Project (OWASP) and Gartner to help CISOs more actively govern one of their most urgent risk categories.
Why agent identity breaks traditional NHI management
Traditional NHI management focuses on discovering an organization's service accounts, bots, and other machine-based identities, then assigning each an owner, granting least privilege, rotating credentials, and monitoring use. It was built around predictable machine activity (e.g., a service account querying a specific database or an API credential connecting two approved applications) where the identity has a stable purpose, an identifiable owner, and a fixed set of permissions.
OWASP's Top 10 for Agentic Applications identifies the risk this creates when AI agents enter the picture. ASI03: Identity & Privilege Abuse occurs when agents inherit or exercise privileges beyond what a specific task requires, and those privileges get reused, escalated, or carried into contexts they were never scoped for.
Here's how that plays out. If CAPTCHA is the bouncer at the door, an AI agent is a VIP host inside the club, responding to several requests at once: “Get my team a table, pull up last month’s tab, send the receipt to our finance contact, and reserve a car home.” Each part of this reasonable-sounding request touches a different system—the guest list, payment records, third-party contacts, a chargeable account—and carries its own data, permission, fraud, and reputational risk. An AI agent with the authority to handle all of it is like a VIP host juggling too many keys.
Traditional NHI management can tell the club that the VIP host has a valid badge and that someone owns it. What it often cannot decide is whether, at that moment, the host should be allowed to:
- View a particular guest's payment data.
- Charge a particular card or exceed a defined amount.
- Use a key borrowed, inherited, or delegated from another staff member.
- Act on a request that was manipulated or embedded in a malicious note.
Traditional NHI controls remain the foundation. But agents need granular, action-time judgment about what should happen right now, not just badge validation.
The governance framework
Gartner's guidance to CISOs on normalizing AI adoption includes classifying agents by autonomy and data sensitivity, enforcing least privilege, and assigning human ownership—one piece of a broader, Zero Trust-oriented push toward collaborative AI governance.
Classify agents by autonomy and data sensitivity
Classifying agents by autonomy means sorting them by how much they can do on their own before a human must intervene. Gartner specifies four levels of AI agent autonomy, shown in the matrix below alongside data sensitivity, since autonomy alone isn't enough to gauge risk.
| Sensitivity → Autonomy ↓ | Public or low- sensitivity data | Internal or confidential data | Restricted, regulated, or critical data |
| Observe Read and retrieve | Low | Moderate | High |
| Advise Recommend or draft | Low | Moderate | High |
| Act with approval Write or communicate after explicit sign-off | Moderate | High | Critical |
| Act autonomously Execute within guardrails | High | Critical | Critical/ redesign |
Increased data sensitivity (e.g. source code, payroll information) paired with higher levels of autonomy calls for a more cautious approach. But it's important to note that "critical/redesign" doesn't mean "never use AI." It means the organization should reduce risk before deployment—for example, by removing standing write privileges in favor of approval gates, or lowering transaction thresholds.
Enforce least privilege
Enforcing least privilege means limiting what actions an agent can take and how long that access lasts. This can look like disabling network access, blocking outside sharing, or cutting off functions at the platform level.
Assign human ownership to every tool
This means identifying whose authority an agent is using to complete its tasks. The distinction matters: an agent may act for an employee, but it shouldn't be indistinguishable from that employee in IAM logs and authorization decisions. I agree with Gartner's recommendation to prohibit the sharing of human credentials and session reuse by agents and to assign every agent a unique identity with an accountable human owner.
A starting checklist for CISOs
- Inventory every non-human identity, including AI agents, not just service accounts and API keys.
- Classify by autonomy and data sensitivity using our matrix framework to flag high-risk combinations.
- Enforce least privilege with time-bound access (i.e. no standing permissions where task-scoped, expiring ones will do).
- Assign a human owner to every agent—no shared credentials, no session reuse.
- Audit regularly, since agent permissions and behavior can drift over time.
To be clear, none of these steps are meant to deter organizations from AI adoption. They just make sure someone's still watching the door.
Frequently asked questions
What is a non-human identity (NHI)?
A non-human identity is any machine-based identity that authenticates or acts within a system: service accounts, API credentials, bots, and increasingly, AI agents. Traditional NHI management focuses on discovering these identities, assigning them an owner, granting least privilege, rotating credentials, and monitoring their use.
Why do AI agents pose a different risk than traditional non-human identities like service accounts?
AI agents pose a different risk than traditional non-human identities because they can inherit or exercise privileges beyond what a specific task requires, and those privileges can be reused, escalated, or carried into contexts they were never scoped for. Traditional NHI management, on the other hand, was built for predictable machine activity, where an identity has a stable purpose, an identifiable owner, and a fixed set of permissions.
What does least privilege mean for AI agents specifically?
Least privilege for AI agents means limiting what actions an agent can take and how long that access lasts. Ways to enforce least privilege include disabling network access, blocking outside sharing, or cutting off functions at the platform level, rather than granting standing permissions.
Why does human ownership matter for AI agents?
Human ownership matters for AI agents because while an agent may act on behalf of an employee, it shouldn't be indistinguishable from that employee in IAM logs and authorization decisions. CISOs can ensure human ownership by prohibiting the sharing of human credentials and session reuse by agents and assigning every agent a unique identity with an accountable human owner.
