Azure penetration testing is an authorised, simulated attack against your own Azure tenant, subscriptions and workloads, used to prove that controls hold up and that alerts actually fire. Microsoft permits this on assets you own once you follow the Microsoft Cloud Unified Penetration Testing Rules of Engagement, and pre-approval is no longer required. Before a single command runs, document written authorisation, list every in-scope asset and switch on full logging.
TL;DR:
- Testing frequency should match your environment's rate of change, with continuous testing for dynamic workloads and annual scans for static setups.
- Scoping must explicitly include tenant, subscription, resource group, tested services, and privilege levels to avoid accidental violations or missed vulnerabilities.
- Use current tools like Microsoft Graph PowerShell SDK and az CLI, running from isolated environments, and throttle requests to prevent Azure abuse flags.
- Verify that detection systems, such as Sentinel or Defender, generate alerts for suspicious activities like role changes or sign-in anomalies, and record whether they fire during tests.
- Reporting should include proof-of-concept, telemetry IDs, and clear severity ratings, with retests confirming issues and alerts are properly configured.
Table of Contents
- Why test Azure: objectives, scope and timing
- Microsoft rules of engagement: what is allowed and what is not
- A staged, assumed-breach methodology worth following
- Tooling and safe automation for Azure engagements
- Proving your monitoring actually catches an attack
- Reporting and remediation that actually close the gap
- How a CREST-accredited team runs this in practice
- What most Azure pentest advice gets backwards
- Azure penetration testing services worth booking
- Sources
- FAQ
Why test Azure: objectives, scope and timing
A pentest exists to answer three questions: do the controls hold under a real attack pattern, does the environment behave the way the architecture diagram says it should, and does anyone notice. Vulnerability scans answer the first question poorly and the third not at all, which is why the Azure Security Benchmark treats red team exercises and scanning as complementary rather than interchangeable.
How often you test should follow your rate of change and threat model, not the calendar. A tenant with frequent deployments and privileged automation needs testing closer to continuous; a static workload with infrequent releases can run a scheduled annual or biannual assessment instead.
Scoping is the part teams rush and later regret. Before testing, you need to define:
- The tenants, subscriptions and resource groups in scope, named explicitly rather than implied.
- The specific services under test (Azure AD/Entra ID, Storage, Functions, Key Vault, VMs) and any excluded production dependencies.
- The identities, roles and service principals testers will use, and what privilege ceiling applies.
Microsoft rules of engagement: what is allowed and what is not
Microsoft's official pentest rules of engagement set clear boundaries. You may test resources you own or have explicit written authorisation for. You may not run denial-of-service style load, you may not test tenants belonging to other organisations, and you may not send phishing campaigns through Microsoft's own messaging infrastructure.
In practice this means:
- Keep signed authorisation and a scope document on hand at all times during the engagement.
- Avoid high-volume fuzzing or brute-force techniques against shared Azure infrastructure, since these resemble denial-of-service traffic.
- Report any genuine vulnerability discovered in Microsoft Online Services itself to MSRC through coordinated disclosure, separately from your client report.
Azure's automated abuse detection sometimes flags legitimate testing as malicious activity. Having authorisation evidence ready lets you respond quickly and avoid an unnecessary account suspension mid-engagement.
Pro Tip: Keep a one-page authorisation summary, dates, scope, contact details, saved locally and accessible offline, in case a flagged account locks you out of cloud-based documentation.
A staged, assumed-breach methodology worth following
Treat the test as assumed breach: assume an attacker already has a foothold and work outward, rather than only probing the perimeter. Practitioner references such as the offensive Azure security cheatsheet organise this into recognisable stages, each producing evidence for your report and for detection validation.
- Reconnaissance: enumerate tenant domains, subscriptions and exposed services without touching production data; record every asset identified.
- Initial access: use provided credentials, a compromised token or a known misconfiguration rather than brute force; log the exact entry vector.
- Authentication and token handling: capture token scopes, refresh behaviour and conditional access responses; note correlation IDs and sign-in event IDs.
- Enumeration: map roles, group memberships, service principals and storage permissions; screenshot or export the raw output.
- Privilege escalation: attempt role assignment abuse or over-permissioned managed identities; document the exact API call and timestamp.
- Lateral movement: move only within authorised subscriptions, never across tenant boundaries; record each hop and the credential used.
- Data exfiltration: simulate access to sensitive data using benign markers or test accounts rather than real records.
- Persistence: place proof-of-possession artefacts, such as a tagged resource or a named test secret, that would trigger the same access path an attacker would use, without exposing genuine data.
Scope discipline matters more here than technique. A single unauthorised step into an adjacent tenant turns a legitimate test into a reportable incident.
Tooling and safe automation for Azure engagements
Legacy identity modules are on their way out and using them is now a liability rather than a convenience. The offensive Azure security cheatsheet flags MSOnline and AzureAD PowerShell as deprecated, in favour of tools built for the current API surface.
- Use the Microsoft Graph PowerShell SDK, Microsoft Entra PowerShell, the Az module and az CLI for identity and resource management tasks.
- Use token-driven tools such as GraphRunner for enumeration and abuse of Graph API permissions once you hold a valid token.
- Run tooling from a controlled context: a dedicated test tenant, an isolated VM using a managed identity, or a separate test subscription, never from a production admin workstation.
Rate-limit scripted calls and log every request your tooling makes locally, independent of Azure's own logs. This protects you two ways: it stops your scanning from tripping Azure's abuse detection, and it gives you an independent record to reconcile against the client's SIEM when you get to detection validation.
Pro Tip: Throttle enumeration scripts to mimic normal API call patterns rather than firing requests as fast as the SDK allows, it reduces false abuse flags without slowing the engagement meaningfully.
Proving your monitoring actually catches an attack
An exploit that succeeds silently has told you nothing useful about your defences. The Well-Architected security testing guidance treats a missed alert on a simulated attack as a detection failure in its own right, equal in weight to the exploit itself.
Certain actions should reliably generate alerts if monitoring is configured correctly:
- Credential theft or unusual sign-in patterns, especially from new locations or devices.
- Role assignment changes, particularly grants of owner or contributor at subscription scope.
- Key Vault secret retrieval events outside normal application access patterns.
Sentinel and Defender rely on analytics rule templates and custom KQL queries to turn raw telemetry into actionable alerts, and testers should record whether the alert fired, how long incident creation took, and whether the retention window covered the event. When an alert fails to fire, the fix is to tune the rule, extend telemetry coverage, or add the missing data source, then retest that specific action in isolation rather than rerunning the whole engagement.
Reporting and remediation that actually close the gap
A report that lists vulnerabilities without proof or priority gets filed and forgotten. A useful Azure pentest report needs an executive summary, the specific impacted assets, proof-of-concept steps, the detection evidence gathered during testing, and a severity rating that maps findings to business impact rather than just a CVSS score.
- Group findings into critical, urgent and standard tiers, with an assigned owner and a target date for each.
- Attach the telemetry identifiers, correlation IDs, event IDs, request IDs, captured during testing so the SOC can trace each finding through its own logs.
- Close a finding only after a non-destructive retest confirms the fix and, where relevant, confirms the alert now fires.
| Severity tier | Typical example | Remediation expectation |
|---|---|---|
| Critical | Over-permissioned managed identity enabling subscription takeover | Fixed and retested before sign-off |
| Urgent | Missing alert on role assignment change | Detection tuned and validated within the agreed window |
| Standard | Verbose error messages exposing internal paths | Scheduled into the next maintenance cycle |
How a CREST-accredited team runs this in practice
Azure penetration testing engagements should follow an assessed methodology rather than an improvised checklist. Authorisation is documented before any testing starts, scope is agreed against the client's actual architecture rather than a generic template, and every finding is retested to confirm the fix holds and that the relevant alert fires. Reports are written in plain English by the tester who ran the engagement, not repackaged by a sales team.

What most Azure pentest advice gets backwards
Most guidance on this topic still reads like a vulnerability scanning checklist with an Azure logo on it: find the misconfigured storage account, flag the over-permissioned service principal, move on. That misses the actual value of a pentest. The point is not the exploit, it is whether your Sentinel workspace or Defender console noticed the exploit happening.

Detection validation gets treated as a nice-to-have when it should be the primary success criterion. A tester who breaches a subscription and reports it as a win, without checking whether an alert fired, has told the client half a story. The reader's priority should be to insist, in the scoping conversation, that every attack path tested comes with a matching detection check, not just a proof-of-concept screenshot.
The second thing conventional advice underweights is tooling hygiene. Scripts built on deprecated AzureAD or MSOnline modules will keep working for a while, then fail silently or behave inconsistently as Microsoft phases them out. Building test tooling on Microsoft Graph PowerShell, Az and az CLI now avoids a rebuild later, and it also produces cleaner, more current telemetry for the SOC to reconcile against.
— Solusec Admin
Azure penetration testing services worth booking
If your Azure environment has never had a proper authorised test, or it has been more than a year since the last one, that is the gap to close first. Solusec runs CREST-accredited penetration testing engagements built around exactly the methodology above: documented authorisation, staged testing, and detection validation rather than a vulnerability list with no proof.

- Penetration testing: a scoped, assumed-breach assessment of your Azure tenant with a plain-English report and retest included.
- Continuous testing: ongoing validation for tenants that change faster than an annual cycle allows, through Solusec's continuous penetration testing service.
- Certification support: once findings are fixed, Cyber Essentials or IASME Cyber Assurance can confirm the baseline is holding.
Engagements should be run by certified testers who communicate directly with clients, not sales contacts relaying findings secondhand. Get in touch through the penetration testing page to scope an Azure assessment.
Sources
FAQ
Does Azure use AES-256 encryption for stored data?
Yes, Azure Storage encrypts data at rest using 256-bit AES in Galois/Counter Mode by default, and this is enabled automatically for all storage accounts. The encryption is FIPS 140-2 compliant, so no manual configuration is needed to meet this baseline.
Is Microsoft Sentinel classified as a SIEM tool?
Yes, Microsoft Sentinel functions as a cloud-native SIEM, using analytics rules and KQL queries to detect suspicious activity across connected data sources. Microsoft has confirmed that after March 31, 2027, Sentinel will be accessed only through the Defender portal rather than as a standalone product.
What are the typical stages of a penetration test?
A staged Azure engagement generally moves through reconnaissance, initial access, authentication and token handling, enumeration, privilege escalation, lateral movement, data exfiltration and persistence, as set out in the offensive Azure security cheatsheet. Each stage should produce its own evidence, such as proof-of-concept steps and telemetry identifiers, rather than a single combined write-up.
Does an Azure SLA guarantee 99.99% availability?
No, an SLA is a contractual commitment with defined exclusions, not a guarantee of that exact availability figure in practice. Microsoft's reliability guidance explains that real resilience comes from architectural choices such as multi-region and availability zone design, with the SLA acting as one input among several.
Do I need Microsoft's permission before testing my own Azure tenant?
No, Microsoft's pen-testing guidance confirms that pre-approval is no longer required for testing resources you own. You must still follow the Microsoft Cloud Unified Penetration Testing Rules of Engagement and keep authorisation evidence ready in case automated abuse detection flags your activity.
Recommended
- Fixed-Price Penetration Testing
- AI Penetration Testing UK | CREST-Accredited
- AI Penetration Testing | CREST-Accredited
This article was created with assistance of AI. We never use AI with your data unless given explicit permission.
