Entra workload identities, UAE

Every control you built protects people. Attackers moved to the identities that are not people.

Microsoft states it plainly: recent cyber attacks show adversaries are increasingly targeting non-human identities over human ones. Applications, service principals and managed identities hold real access, hold multiple credentials, and almost nobody reviews them.

Entra workload identities for UAE organisations
  • 3 typesApplications, service principals, managed identities
  • No MFABecause there is nobody to prompt
  • Multiple credsPer workload, unlike a human user
  • Rarely revokedBecause nobody knows when they should be
What we assess and control

Eight things that govern the non-human half of your directory.

Workload identities are the population where the usual controls do not apply. There is no MFA to require, no user to educate, no leaver process to trigger. What is left is inventory, credential hygiene, policy and review, and most estates do none of the four.

Knowing what you actually have

In Entra, workload identities are applications, service principals and managed identities. An application object is the global representation across all tenants, and a service principal is the local instance in a specific tenant that defines what the app can actually do there. The inventory almost always surprises people.

Credential sprawl, which is structural rather than careless

Microsoft puts it well: a human user typically has a single identity used across a broad range of resources, while a software workload may deal with multiple credentials for different resources, all of which need to be stored securely. That asymmetry is why credential hygiene is harder here.

Conditional Access for service principals you own

Conditional Access policies can be applied to service principals owned by your organisation, which brings location and risk conditions to non-human sign-ins. The scoping detail matters: this covers service principals your organisation owns, not every third-party application present in the tenant.

Real-time enforcement through continuous access evaluation

Continuous access evaluation for workload identities enables real-time enforcement of Conditional Access location and risk policies. That closes the same gap it closes for people: the window between a condition changing and the session actually stopping, which for a workload can otherwise be a long time.

Risk detection, including leaked credentials

Entra ID Protection detects risks such as leaked credentials for workload identities, contains threats and reduces risk. Leaked credentials are the detection that matters most here, because a service principal secret committed to a repository is one of the most common ways this population gets compromised.

Removing secrets entirely where you can

Managed identities are a special type of service principal that eliminates the need for developers to manage credentials, for workloads running on Azure. Workload identity federation covers the rest, for supported scenarios such as GitHub Actions, workloads on Kubernetes, and compute platforms outside Azure.

Reviewing the privileged ones

Access reviews for service principals cover service principals and applications assigned to privileged directory roles. This is the review nobody runs, and it is the one that finds the application granted a directory role during a project three years ago that nothing has used since.

Agent identities, the newest population

Entra Agent ID provides identity constructs for AI agents, offering enforced human sponsorship, lifecycle governance from provisioning through deactivation, and at-scale management applying centralised policy across all agent instances of a type. Worth designing for now rather than retrofitting later.

Why this population is different

It is hard to track when a workload identity is created, or when it should be revoked.

Microsoft names that specific difficulty, and it explains why non-human identities accumulate in a way human accounts do not.

  • A human account has a creation event somebody requested and a revocation event somebody eventually triggers, because a person joined and a person left. A service principal is created by a developer, a deployment, or a consent prompt, and nothing about the organisation changes when the thing it served stops being used.
  • The credential position compounds it. A human user typically has one identity across many resources. A software workload may hold multiple credentials for different resources, each needing secure storage, each with its own expiry that somebody has to notice.
  • That is why the assessment starts with inventory rather than policy. You cannot apply Conditional Access to service principals owned by your organisation, or review the ones holding privileged directory roles, until you know which exist and what each was for.
  • The direction of travel is to remove credentials rather than manage them. Managed identities eliminate credential management for workloads running on Azure, and workload identity federation covers supported scenarios outside it including GitHub Actions, Kubernetes and other compute platforms. Every workload moved onto those is one fewer secret to rotate.
Ask us to inventory your service principals
How we approach it

Four things that make this a programme rather than a clean-up.

Deleting unused service principals feels productive and changes nothing structurally. What changes the position is ownership, credential elimination and a review that recurs.

We attribute ownership before we change anything

Microsoft notes it is hard to track when a workload identity is created or when it should be revoked. Attribution is the work that fixes that, and it is conversations rather than queries. An identity with a named owner can be reviewed, reduced and eventually retired. One without cannot be touched safely.

We remove secrets rather than rotating them better

Managed identities eliminate the need to manage credentials for workloads running on Azure, and workload identity federation covers supported scenarios such as GitHub Actions, Kubernetes and compute outside Azure. Every workload moved is one fewer secret that can leak or expire at an inconvenient moment.

We apply policy where it genuinely applies

Conditional Access for workload identities covers service principals owned by your organisation, and continuous access evaluation adds real-time enforcement of location and risk. Being precise about that scope avoids the common assumption that a policy covers every application present in the tenant.

We start the review with the privileged subset

Access reviews for service principals cover applications assigned to privileged directory roles, and that subset is where the risk concentrates. Reviewing a hundred privileged workload identities properly is better use of effort than skimming several thousand, and it is where the uncomfortable findings are.

How an engagement runs

Four phases across roughly six to eight weeks.

Inventory dominates the early phases because the population is genuinely unknown in most tenants, and because attributing each identity to an owner takes conversations rather than queries.
  1. 01
    Weeks 1 to 2

    Inventory and attribute ownership

    Every application, service principal and managed identity, with what it can access, what credentials it holds and when they expire. Then the harder half: who owns each one and what it is for. Identities nobody can attribute are the finding, not an inconvenience.

    • Full workload identity inventory
    • Permissions and directory roles mapped per identity
    • Credential inventory with expiry dates
    • Unattributable identities listed separately
  2. 02
    Weeks 3 to 4

    Reduce permissions and remove secrets

    Over-permissioned identities brought down to what they use, and credentials eliminated where the workload can move to a managed identity or to workload identity federation. Every secret removed is one that cannot leak, expire unexpectedly, or need rotating.

    • Excess permissions identified and reduced
    • Managed identity migration candidates confirmed
    • Federation candidates identified for GitHub Actions and Kubernetes
    • Remaining secrets moved to proper storage with rotation
  3. 03
    Weeks 5 to 6

    Apply policy and detection

    Conditional Access applied to service principals your organisation owns, continuous access evaluation enabled for real-time location and risk enforcement, and ID Protection risk detection reviewed so leaked credential findings reach somebody who will act on them.

    • Conditional Access policies for owned service principals
    • Continuous access evaluation enabled where supported
    • Risk detection reviewed and routed to an owner
    • Custom security attributes applied for classification
  4. 04
    Weeks 7 to 8

    Establish review and lifecycle

    Access reviews configured for service principals and applications in privileged directory roles, a creation process that captures ownership at the point of creation, and a decommission path so identities are removed when what they served is retired.

    • Access reviews configured for privileged service principals
    • Creation process capturing owner and purpose
    • Decommission path defined and tested
    • Agent identity approach agreed if AI workloads are in scope
Where this matters

Six situations where non-human identity is the exposure.

The common feature is an organisation that has done the human identity work well and has never looked at the other half of the directory.

A business that found a secret in source control

The most common trigger, and Entra ID Protection detects leaked credentials for workload identities specifically. The immediate response is rotation, and the useful response is asking how many other secrets exist, where they are, and which of those workloads could stop using secrets entirely.

A regulated firm asked to evidence privileged access review

The review usually covers people and stops there. Applications and service principals assigned to privileged directory roles hold comparable access and are almost never included. Access reviews for service principals close that gap and produce the evidence in the same form as the human review.

An organisation with years of accumulated integrations

Every project leaves a service principal, and nothing removes them because nothing about the organisation changes when a project ends. Inventory and attribution typically find a meaningful proportion that nobody can account for, which is both a risk finding and a licensing conversation.

A company running CI/CD pipelines into Azure

Pipelines authenticating with long-lived secrets are a well understood risk and an avoidable one. Workload identity federation is supported for scenarios including GitHub Actions and workloads running on Kubernetes, which removes the stored credential from the pipeline entirely.

An operator whose outage was a credential expiry

A client secret expires, an integration stops, and nobody knew the date because nobody owned the identity. That is a availability incident caused by an identity governance gap, and the fix is inventory with expiry tracking, or better, moving the workload off secrets entirely.

A business starting to deploy AI agents

Agent identities are a distinct category with their own governance requirements, and Entra Agent ID provides enforced human sponsorship, lifecycle governance from provisioning through deactivation, and centralised policy across agent instances. Designing that in now is considerably easier than retrofitting it.

Three positions

How UAE organisations govern non-human identity.

The right column is the norm, and the middle column is where most organisations land after a single incident. The left column is a programme rather than a project.
Complete inventory
Governed as a populationMaintained
Reactive after an incidentPoint in time
UngovernedNone
Ownership attributable
Governed as a populationPer identity
Reactive after an incidentFor some
UngovernedNo
Permissions right sized
Governed as a populationYes
Reactive after an incidentFor the ones examined
UngovernedNo
Secrets eliminated where possible
Governed as a populationManaged identity and federation
Reactive after an incidentPartly
UngovernedNo
Conditional Access applied
Governed as a populationTo owned service principals
Reactive after an incidentNo
UngovernedNo
Real-time enforcement
Governed as a populationContinuous access evaluation
Reactive after an incidentNo
UngovernedNo
Leaked credential detection acted on
Governed as a populationRouted to an owner
Reactive after an incidentSometimes
UngovernedUnseen
Privileged ones reviewed
Governed as a populationAccess reviews
Reactive after an incidentNo
UngovernedNo
Decommission path exists
Governed as a populationYes
Reactive after an incidentNo
UngovernedNo
AI agent identities planned for
Governed as a populationYes
Reactive after an incidentNo
UngovernedNo
Human against non-human

Why the controls you already built do not transfer.

Each row is a control that works well for people and either does not exist or does not apply for workloads. That gap is the reason this population needs its own programme.
ControlHuman identitiesWorkload identities
Multifactor authenticationStandardNot applicable, nobody to prompt
Number of credentialsTypically one identityPotentially many, per resource
Creation triggerA person joinsA developer, deployment or consent
Revocation triggerA person leavesFrequently nothing at all
Conditional AccessFull policy setService principals your organisation owns
Continuous access evaluationSupportedSupported for workload identities
Risk detectionID ProtectionID Protection, including leaked credentials
Periodic reviewAccess reviewsAccess reviews for service principals in privileged roles
Credential eliminationPasswordless methodsManaged identities and workload identity federation
Security awareness trainingApplicableNot applicable
How an engagement runs

Five steps, and the first one takes longer than it should.

Inventory is quick. Attributing each identity to an owner and a purpose is what takes the time, and it is what makes everything afterwards possible.
  1. 1

    Inventory applications, service principals and managed identities

    What exists, what each can access, which hold privileged directory roles, which are yours and which belong to third parties. The distinction between your own service principals and third-party ones matters because Conditional Access for workload identities applies to the ones your organisation owns.

  2. 2

    Attribute ownership and purpose

    A named owner and a stated purpose per identity. This is the step that converts a list into something actionable, and identities nobody can attribute get flagged rather than quietly deleted, because deleting an unattributed identity is how an outage happens.

  3. 3

    Reduce permissions and eliminate credentials

    Permissions brought down to what is actually used, and workloads moved onto managed identities where they run on Azure or workload identity federation where they run elsewhere. Remaining secrets moved into proper storage with tracked expiry and a rotation owner.

  4. 4

    Apply policy and route the detections

    Conditional Access on service principals your organisation owns, continuous access evaluation for real-time location and risk enforcement, and ID Protection workload identity risk detections routed to somebody who will act, particularly leaked credential findings.

  5. 5

    Establish review and lifecycle

    Access reviews for service principals and applications in privileged directory roles, a creation process that captures ownership at the point of creation rather than years later, and a decommission path so retirement of a system retires its identity too.

Straight answers

What organisations ask about workload identities.

Self assessment

Fifteen questions about the non-human half of your tenant.

If you cannot answer the first three, that is the engagement. Very few organisations can, and it is not a reflection on the team.

Inventory

  • How many service principals exist?
    A number, not an impression.
  • How many hold a privileged directory role?
    This is the critical subset.
  • Can we name an owner for each?
    Usually not.
  • How many managed identities do we run?
    And what do they access.
  • Which are third party rather than ours?
    Policy scoping depends on it.

Credentials

  • How many client secrets are in use?
    And where are they stored.
  • When does the next one expire?
    Expiry causes outages.
  • Has any secret ever been in source control?
    Check, do not assume.
  • Could this workload use a managed identity?
    It eliminates credential management.
  • Could this pipeline use federation?
    GitHub Actions is supported.

Control

  • Any Conditional Access on service principals?
    For ones we own.
  • Is continuous access evaluation enabled?
    For real-time enforcement.
  • Who sees workload identity risk detections?
    Leaked credentials especially.
  • Do we review privileged service principals?
    Access reviews support this.
  • Do we have AI agents in the tenant?
    Agent ID governs them.
Next step

Count your service principals, then count the ones with a privileged directory role.

Tell us about application registrations, service identities, credential records and integration owners. We will confirm the requirements and proposed scope with you.