What a PAA Azure Security Assessment Found: Anonymised Findings from a Dutch Scale-Up Environment

An anonymised walkthrough of PAA assessment findings from a Dutch scale-up environment: 85 employees, 3 subscriptions, no prior formal assessment. The findings are specific. The NIS2 gaps are real.

Marc Dekeyser |

What a PAA Azure Security Assessment Found: Anonymised Findings from a Dutch Scale-Up Environment

Six of the eleven engineers at this company held permanent Owner-level access to at least one production subscription.

Not elevated. Not time-limited. Permanent. Always on.

The security team knew it was a problem. What they didn’t know was the number. They guessed two.

A structured PAA assessment of their 85-person Azure tenant surfaced 34 findings: 6 critical, 11 high, spread across identity, network, and logging. Almost nothing was exotic. Most of it was configuration drift the team couldn’t see, because nobody had ever looked at the whole tenant at once.

That gap, between what the team believed and what was actually running, is where the findings that matter lived. Not sophisticated attack paths.

Drift.

I kept the finding IDs and the severities because the specificity is the whole point. The environment is anonymised, the numbers aren’t. This is what “drift nobody had a clear picture of” looks like, line by line.


The environment

Dutch SaaS provider. 85 employees. An important entity under NIS2 in the digital infrastructure sector. Azure-native since day one, no on-premises, no hybrid connectivity. The platform team is four engineers.

Production runs across three subscriptions: sub-prod-platform, sub-prod-customer, and sub-prod-data. A fourth, sub-dev, carries non-production. Above subscription level there’s nothing. No management group structure, the subscriptions sitting directly under the root tenant management group with no intermediate groupings. No prior Well-Architected assessment, ever.

PAA connected through a read-only service principal with Reader and Security Reader at management group scope. Azure Resource Graph queries ran across all four subscriptions. Including the NIS2 control mapping, the whole thing took about four hours.

What follows is a representative slice of the full report: the security pillar and the NIS2 mapping module.


Identity and access management findings

IAM-001 — Critical Privileged Identity Management not activated for 6 privileged role assignments

PIM is enabled in the tenant. It just isn’t applied to six role assignments at subscription scope. Three are Owner on sub-prod-customer, two are Owner on sub-prod-data, one is Contributor on sub-prod-platform. All six are direct user principal assignments, not group-based. All six are active assignments, not eligible ones waiting on activation.

So any of those six principals, or anyone who has taken over one of their accounts, has permanent unrestricted access to production. No activation event. No justification. No audit trail that separates a legitimate admin action from an attacker moving sideways. The Azure Activity Log records what was done. It cannot tell you whether the person doing it was the account holder.

“But our engineers need Owner. We’re a small team, we can’t have people waiting on approvals at 2am.”

Nobody is asking them to wait. Convert all six to eligible assignments through PIM, require an activation justification, cap the activation window at four hours, and turn on PIM Access Reviews every 30 days. Self-activation takes seconds. What changes isn’t the access, it’s that there’s now a record of someone reaching for it.

And PIM is already in this tenant. It ships with the Microsoft Entra ID P2 licences the company is already paying for.

This one costs nothing to close.

IAM-002 — High 14 service principals with active client secrets, 4 with secrets not rotated in over 365 days

Twenty-two service principals carry application role assignments. Fourteen of them authenticate with client secrets rather than certificates or federated credentials. Four of those fourteen have secrets last modified more than 365 days ago. One secret hasn’t been rotated in 847 days.

Three of the four long-lived secrets belong to principals with Contributor-level access on production subscriptions. The fourth holds Key Vault Secrets User on two Key Vault instances in sub-prod-data.

This isn’t negligence, and I want to be clear about that. Manual secret rotation needs the team that owns the principal and the team that consumes it to coordinate. It creates deployment risk. There’s no automated enforcement anywhere. The rotation gap is what the system produces, not what the people did wrong.

The way out is Managed Identity for runtime workloads, which kills the rotation problem outright, and Workload Identity Federation for CI/CD pipeline authentication.

IAM-003 — Medium No Conditional Access policy enforcing MFA for guest access to Azure Portal

Eleven guest accounts have access to Azure resources here. The Conditional Access policy that enforces MFA covers members only. Guests are explicitly out of scope. Three of the eleven guests hold at least Reader access on production subscriptions.

One Conditional Access policy, scoped to guest users, requiring MFA and trusted-device compliance for Azure Portal and Azure Resource Manager access, closes the gap. Internal member access stays untouched.


Network and data protection findings

NET-001 — Critical Inbound NSG rules allowing TCP/22 from 0.0.0.0/0 on 2 subnets

Two Network Security Groups allow SSH from any source address. Priority 100 in both. No IP restriction at all.

One of them, nsg-dev-jumpbox, sits on subnet snet-management in sub-dev. Development subscription, lower immediate risk. The other is nsg-prod-legacy, on subnet snet-app-tier in sub-prod-platform. That one is not development. The subnet holds three Virtual Machines. The rule was created 14 months ago and nobody has touched it since.

I see this constantly. A rule gets opened for a reason that made sense at the time (someone needed SSH during an incident, or a deployment wanted it for an afternoon) and then it just stays. The immediate need passes. The rule doesn’t. I have left one of these open myself, which is probably why I spot them so fast now.

Remove the 0.0.0.0/0 inbound SSH rules. If SSH access is genuinely required, route it through Azure Bastion. Standard tier is already deployed in the hub resource group and peering is active, so this is a routing configuration change, not a new deployment. If direct SSH from specific IPs is a hard requirement, restrict to those ranges and write the justification into the NSG rule description field so the next person knows why it’s there.

NET-002 — High Storage Account sa-prod-backups does not require secure transfer and allows public blob access

sa-prod-backups in sub-prod-data has supportsHttpsTrafficOnly set to false and allowBlobPublicAccess set to true. It holds the platform tier’s backup archives on a 30-day retention policy.

Secure transfer is a single property change. The public-access flag is more nuanced. It being enabled doesn’t automatically make containers publicly accessible (container-level ACLs govern actual access), but it strips out a control layer and it lights up audit findings under both the Azure Security Benchmark and Microsoft Defender for Storage. Walk each container. For backup content, no container should carry anonymous read access. Then disable public access at the account level once you’ve confirmed nothing depends on it.

NET-003 — Medium Microsoft Defender for Storage not enabled on 3 storage accounts in sub-prod-data

sa-prod-backups, sa-data-lake-01, and sa-customer-exports, all in sub-prod-data, are missing Microsoft Defender for Storage. Every other storage account in scope has it through the subscription-level Defender for Cloud plan.

Why these three? They were created after the subscription-level plan was switched on, and the plan doesn’t reach back to resources that carry existing exclusion configurations. Turning Defender on for all three runs about €14/month combined at current transaction volume. Leaving it off means missing detection coverage for malware upload attempts, anomalous access patterns, and sensitive data exposure events.

Fourteen euro.

LOG-001 — High Log Analytics workspace retention set to 30 days; Microsoft Sentinel not deployed

law-platform-centralised in sub-prod-platform keeps data for 30 days. That’s the Azure default. It’s also not enough for regulatory compliance.

NIS2 Article 21(2)(b) wants measures that include the ability to detect and respond to incidents. A 30-day window makes post-incident forensic analysis impossible for anything you didn’t catch inside that month. ISO 27001 Annex A controls 8.15 and 5.28 both point at retention long enough to actually investigate. Ninety days is the usual floor. Log Analytics Interactive Retention stretches to 90 days with no change in query cost, so this is a setting, not a line item.

Sentinel isn’t deployed. The platform team runs manual alert rules in Azure Monitor. I won’t pretend Sentinel is mandatory for a minimum viable security operation, because it isn’t. But without it there’s no correlation across identity events, resource configuration changes, and network anomalies. Detection stays point-in-time instead of pattern-based.


Where the NIS2 gaps landed

Three Article 21 obligations came out with material gaps in this environment.

Incident handling, 21(2)(b). The 30-day log retention and the absent Sentinel are a direct hit. You cannot detect, analyse, and recover from an incident you can’t reconstruct, and 30 days doesn’t give you the reconstruction.

Authentication and access control, 21(2)(i). IAM-001, IAM-002, and IAM-003 all land here. Six permanent privileged assignments with no PIM. Fourteen service principals on client secrets. Guests with no MFA. Individually, each is a gap against the requirement for multi-factor authentication and secure access control. Together they mean one compromised privileged account, or one leaked Contributor secret, buys an attacker unrestricted production access with no time limit and no activation gate. The blast radius is the entire production environment.

Security in acquisition, development and maintenance, 21(2)(e). NET-001 and NET-002 sit here. An open SSH rule on a production subnet and a storage account that doesn’t enforce secure transfer both fail a basic security-in-maintenance review. Neither is clever. They’re what happens when nobody revisits a configuration after the reason for it expired.


What the whole thing added up to

The full PAA assessment for this environment produced 34 findings across the Security, Operational Excellence, and Reliability pillars: 6 critical, 11 high, 14 medium, 3 low. The NIS2 module flagged 9 controls with partial or no coverage. The complete report, remediation guidance included, was ready within four hours of connecting the service principal.

None of it was clever. That’s the part worth sitting with. There was no novel attack path, no zero-day, no misconfigured thing you’d need a specialist to spot. There was a tenant nobody had looked at end to end, and a pile of small reasonable decisions that nobody went back to unwind.

PAA is the tool I ran this with. I built it because I was tired of doing this by hand: clicking through blades for days, filling in a spreadsheet nobody would open twice, and sending a report that was already stale by the time it landed. The four hours above is the entire reason it exists.

That’s the pitch. It’s the only one you’ll get from me.

If this sounds like your environment (a handful of subscriptions, a small platform team, no prior formal assessment, a compliance obligation getting closer) then here’s the uncomfortable question worth answering before an auditor or an incident answers it for you. How many people hold permanent Owner access to your production subscriptions right now? Not your best guess. The actual number.

If you’re not certain, and most teams aren’t, that uncertainty is the finding.


FAQ

How long does a PAA Azure security assessment take?

For this 85-person, three-subscription environment, the full assessment including NIS2 control mapping finished in roughly four hours, starting from the moment a read-only service principal with Reader and Security Reader roles connected. Azure Resource Graph queries ran across all four subscriptions. No agents were installed.

What were the most serious findings?

Six privileged role assignments with no Privileged Identity Management, granting permanent always-on access to production; an inbound NSG rule allowing SSH from 0.0.0.0/0 on a production subnet; and a storage account holding backups with secure transfer disabled and public blob access enabled. Six findings in total were rated critical.

How do the findings map to NIS2?

Three NIS2 Article 21 obligations had material gaps. Article 21(2)(b) incident handling, from 30-day log retention and no Microsoft Sentinel; Article 21(2)(i) authentication and access control, from the privileged-access and service-principal findings; and Article 21(2)(e) security in development and maintenance, from the open SSH rule and storage misconfiguration.

Do these gaps mean the team was negligent?

No. Most of these are what Azure produces by default, not how the people behave. Manual secret rotation carries deployment risk and has no automated enforcement. NSG rules opened during an incident never get closed. The value of an assessment is seeing the whole tenant at once, which nobody here had done.

Want a second pair of eyes on it?

Ask for a free second opinion