Active Directory sits underneath nearly everything an enterprise runs, authenticating users, enforcing group policy, and gating access to systems most staff never think about. Because it works quietly, teams assume it works correctly.
Environments degrade gradually instead. Staff turnover, ad hoc permission grants, incomplete decommissioning, and configuration drift accumulate over years until something fails or somebody hostile finds the gap first.
One assumption underpinned the traditional review cadence: drift moves at human speed. Somebody joins, changes roles, leaves, and a semiannual check catches the residue.
That assumption is under pressure. Non-human identities, meaning service accounts, API keys, OAuth tokens, machine certificates, and the credentials AI agents hold, now outnumber human users by roughly 45 to 1 on average, according to Cloud Security Alliance research, with cloud-native environments reaching 144 to 1, up from 92 to 1 in the first half of 2024. Palo Alto Networks puts the average at 109 machine identities per human.
Those figures range widely because they count different things in different environments, and ratios in this space usually reflect methodology as much as reality. The direction is not in dispute. The population your directory governs is mostly not people, and it is growing faster than the review schedule was designed for.
How Do Misconfigurations Accumulate?
Gradually, through legitimate changes made under time pressure.
Few environments are misconfigured all at once. A temporary permission granted for a project and never revoked. A group policy object built to solve a problem nobody remembers. A domain controller configured slightly differently because whoever built it followed a different internal guide.
Individually, these rarely cause harm. Collectively, they produce an environment that resists auditing and confuses new staff.
That pattern is why periodic review beats a one-time cleanup. An environment corrected once drifts again within months, since subsequent changes rarely get the scrutiny the original build received. A structured Active Directory health check treats configuration review as an ongoing practice rather than a project with an end date, which matches how these environments actually behave.
Why Has Privilege Sprawl Gotten Worse?
Because the population creating it changed character.
Human privilege sprawl follows a familiar shape. Employees move between roles or take temporary projects, and access outlives its purpose. An account that needed elevation for a six-month project still holds it years later. Multiply across thousands of accounts and a substantial share of privileges no longer map to any job.
Health checks surface that directly, comparing group memberships and delegated permissions against what active roles require. Findings usually run longer than IT teams expect, particularly around privileged groups like Domain Admins, where membership grows through convenience rather than decision.
Non-human identities break the pattern in three ways that matter for review.
They never leave. No offboarding workflow fires, no manager notices an empty desk, and nothing triggers a revocation. GitGuardian’s State of Secrets Sprawl 2026 found 64% of secrets confirmed valid back in 2022 still unrevoked by January 2026, which is four years of exposure on credentials that should have rotated within hours.
They start over-permissioned. Teams grant broad access during testing so an agent never hits a permissions wall, then ship to production with those permissions intact. The 2026 Infrastructure Identity Survey found 70% of organisations granting AI systems more access than they would give a human doing the same job. Sysdig’s data suggests both human and machine identities actually use around 2% of what they have been granted.
They behave differently under monitoring. Machine identities bypass MFA by design, run continuously without working hours, and produce no familiar location, device, or timing signals. Detection tuned to human compromise patterns has little to work with.
What Should a Review Ask About Agents?
Not who has access. What has access to what.
Most organizations can explain what their AI agents are for. Far fewer can state what those agents can reach, how that access is bounded, when permissions get revoked, or which systems inherit that access downstream.
The gap has a governance shape rather than a technical one. Agents acquire permissions at runtime, spawn sub-agents, call external APIs, and chain actions across many systems, which extends the blast radius of one compromised credential well past what a static service account could reach. The Salesloft-Drift incident showed the mechanism, with compromised OAuth tokens letting attackers move across trusted connections into hundreds of downstream organizations.
Questions worth adding to a standard review:
- Which service accounts and OAuth grants exist, who owns each, and what justified the scope originally
- Which of them have not authenticated in ninety days
- Which hold membership in privileged groups, and whether any still needs it
- Whether credentials rotate on a defined schedule or persist indefinitely
- Whether any agent operates with permanent privileged access where just-in-time would serve
Agent activity also generates traffic patterns that older network security assumptions do not anticipate, so a clean monitoring dashboard is weaker evidence than it used to be.
Provisioning frequently happens outside IT entirely. Purpose-built AI tools now cover narrow tasks the general platforms handled poorly, and a department head connecting one to a business system creates a directory principal nobody catalogued. Discovery belongs in the review, not just assessment of known accounts.
Why Do Replication Problems Go Unnoticed?
Because they degrade rather than fail.
Active Directory depends on replication to keep domain controllers synchronized. A controller that has fallen behind serves slightly stale group policy or authentication data, producing inconsistent behavior that affects some users and not others.
Teams often spend real time troubleshooting what presents as an isolated application fault before finding a replication failure that had been accumulating errors for weeks.
A structured check reviews replication status across every domain controller, flagging failures, latency, and topology problems that daily monitoring misses. Because these worsen gradually, waiting for a visible symptom means the underlying issue has usually been present far longer than anyone realized.
What Does Resilience Assessment Cover?
How the environment holds up under stress, whether from hardware failure, ransomware, or a routine change gone wrong.
Backup coverage deserves specific attention. Many organizations back up servers broadly without verifying that Active Directory recovery procedures have been tested. A backup never restored in practice provides confidence rather than resilience.
Assessment also examines single points of failure in domain controller architecture, redundancy across sites, and whether documented procedures describe the environment as it exists now rather than as it looked when someone wrote the runbook.
Four areas consistently repay attention:
- Backup and restore procedures tested against realistic recovery scenarios
- Domain controller placement and redundancy across physical or geographic sites
- Time synchronization accuracy, since drift causes authentication failures
- Documentation currency, so recovery runbooks match actual configuration
Gaps here rarely disrupt daily operations, which is precisely why they survive until an incident forces attention.
How Often Should Health Checks Happen?
More often than the population being reviewed doubles.
A single assessment gives a snapshot, and environments keep changing after it. Annual or semiannual cadence caught human-paced drift reasonably well, since staff turnover and role changes move slowly.
That calculation shifts when machine identities expand faster than human oversight. Palo Alto Networks reported organizations expecting 85% growth in AI agents over twelve months, with machine identities projected to rise 77%. A review model built around periodic checks and long-lived secrets struggles against a population growing at that rate.
Cadence also builds institutional memory. A team that knows what normal looks like for their environment spots meaningful deviation quickly, while one encountering it fresh each time starts from zero.
Organizations handling agentic deployments well build fixed review points into the workflow rather than granting standing autonomy, the approach government agencies took with their agentic AI rollouts. Directory review is where that principle becomes concrete.
One organizational note worth flagging. Executives frequently believe least privilege is enforced because they are looking at human access, while practitioners working with machine and automated systems report otherwise. A health check that reports on both populations separately resolves that disagreement with evidence rather than opinion.
FAQs
Q. What does an Active Directory health check cover?
Configuration drift, privilege sprawl, replication health, and resilience gaps including backup testing, redundancy, and documentation accuracy.
Q. Why do machine identities matter for AD review?
They now substantially outnumber human accounts, they never trigger offboarding, and they frequently hold more access than the equivalent human role would.
Q. How often should reviews run?
Annual or semiannual has been the norm. Environments deploying AI agents at pace should consider shorter intervals, since the identity population grows faster than that cycle.
Q. What is the most commonly overlooked finding?
Untested backups. Broad server backup is common; verified Active Directory recovery is not.
Q. Do standard monitoring tools catch these problems?
Partially. Replication degradation and non-human credential misuse both tend to sit below the threshold that routine monitoring flags.
Q. Where should a team start if they have never done one?
Inventory and privileged group membership. Knowing what exists and what holds elevation frames every other finding.
The Bottom Line
Active Directory rarely fails without warning. It degrades through accumulated misconfiguration, unreviewed privilege, quiet replication faults, and resilience gaps that only surface during an incident.
Structured review addresses that pattern directly, catching problems while they remain isolated. What changed is the composition of what gets reviewed. A directory built around employees now governs a population mostly made of service accounts, tokens, and agents that hold broad permissions, never log off, and leave none of the behavioral traces detection tools were tuned to find.
For infrastructure this central, recurring scrutiny is not caution. It matches how much depends on it, and how quickly the thing being scrutinized keeps growing.
Related: Why Data Diversity Matters for Machine Learning Accuracy
