top of page

Pre-Standard Isn't Optional: Scoring Your AI Agents Against NIST's 10 Identity Tenets

Kostas Tsiolas
Nov 7
4 min read

NIST says the tenets are a starting point. Microsoft has already shipped one as a default. Here's how to find out where you stand before a vendor or an auditor decides for you.


Two events, eight days apart

On 27 October 2026, NIST published ten design tenets for agent identity management systems. They are the second post in a series from NIST researchers and MITRE, and they will guide the agentic demonstrations at NIST's National Cybersecurity Center of Excellence (NCCoE).

NIST positions them as an initial set. A starting point. Not a formal standard, and not a certification bar.

On 4 November 2026, Microsoft Entra began rolling out a change (MC1506728). New agent blueprints no longer receive the AgentIdentity.CreateAsManager permission automatically. Until now, that grant let agents create further agent identities with no administrator approval. From here, an admin has to assign it explicitly, which makes it visible, auditable and revocable.

That is NIST's third tenet, no standing privilege, shipped as a platform default.


The mistake: waiting for the standard

The most common response I hear from clients on AI agents is: “We'll decide once there's a clear standard.” It sounds prudent. It isn't.

While you wait, vendors set defaults, agents keep appearing inside tools you already license, and auditors form expectations. Pre-standard doesn't mean optional. It means you get to set your own baseline, if you set it first.


Most organisations can't see what they'd be governing

The Cloud Security Alliance surveyed 418 IT and security professionals in January 2026 for Autonomous but Not Controlled. 68% rate their visibility into AI agents as high, yet 82% discovered at least one unknown agent in the past year. Only 21% have a formal process to decommission agents, and only 11% automatically block an agent action that crosses a boundary.

If you cannot answer how many agent identities you have and who created them, the other nine tenets have nothing to stand on.


What Microsoft's change means if you run Entra Agent ID

The change affects existing agents as well as new ones where they depend on the automatic grant. Existing identities and blueprints are not removed. Agents that create identities through a directory role are not affected.

Before enforcement reaches your tenant, review Entra audit logs under ApplicationManagement and Add Service Principal, identify agent blueprints that rely on the automatic grant, decide whether each agent should create other agents, and update runbooks because unremediated requests will fail with a 403.

Dates refer to Microsoft's worldwide rollout, expected to complete by late November 2026. Confirm timing in your tenant's Message Center. Also review the tenant setting that lets regular users register agent identity blueprints, which remains on by default.


The self-assessment: ten questions, one per tenet

Score each question honestly: 2 points means yes with evidence today, 1 means partly or only for some agents or platforms, and 0 means no or don't know.

1. Distinct agent identification: Can we produce a complete list of agent identities, separate from human users and service accounts, with who created each one?

2. Linking, delegation and attenuation: Can we trace whose authority an agent acts on, and confirm its rights can only narrow, never widen, as it delegates?

3. Least and zero standing privilege: When an agent is idle, does it hold no rights at all?

4. Short-lived, revocable authorization: Are credentials short-lived, and can we revoke one agent's access in minutes without breaking others?

5. Human-in-the-loop triggers: Which actions require human approval, and are approvals genuinely reviewed?

6. Mandates and constraints: Does each agent have a documented purpose and limits that downstream systems can check?

7. Separating policy from reasoning: Is the system deciding what an agent may do separate from the model deciding what it tries to do?

8. Deterministic access decisions: Are access decisions made by deterministic rules, not by the agent's own judgment or another model?

9. Continuous monitoring: Do we monitor agent activity with agent-specific baselines, not just human user analytics?

10. Immutable evidence: Could we prove from tamper-evident logs held outside the agent what it did last Tuesday and on whose behalf?


Reading your score

0–7: Foundations missing. Start with questions 1, 3 and 10.

8–14: Partial control. Close the inventory gap across platforms first.

15–20: Defensible baseline. Focus on tenets 5, 6 and 7.

Be strict with the 2s. “We think so” is a 1.


Where the regulatory pressure sits

NIS2 and DORA both push organisations from periodic access reviews toward continuous, evidenced control. Agents hold access. An agent estate you can't inventory or evidence is an access-control gap under either regime.


Where to start this month

1. Count every agent identity and platform, with a named creator.

2. Own each agent through one accountable human, not a team.

3. Prove that activity is logged somewhere the agent can't alter, and test that you can retrieve it.


Conclusion

NIST's tenets will change as the NCCoE demonstrations progress. Microsoft will ship more defaults. Neither will wait for your organisation to be ready.

The organisations that come out of this well won't be the ones that waited for the final standard. They'll be the ones that know how many agents they run, who owns each one, and what the evidence says.

If you scored below 15, or couldn't answer question 1, that is the conversation to have now. Nimbus Cyber helps CISOs and CIOs assess identity governance for human, non-human and AI agent identities, and turn the gaps into a prioritised roadmap. Talk to us about an assessment:Book a Diagnostic


Sources

NIST Cybersecurity Insights, Design Tenets for Agent Identity Management Systems: https://www.nist.gov/blogs/cybersecurity-insights/design-tenets-agent-identity-management-systems

NCCoE, Software and AI Agent Identity and Authorization project: https://pages.nist.gov/nccoe-ai-identity/

Microsoft Entra Message Center, MC1506728: https://mc.merill.net/message/MC1506728


Comments


bottom of page