The Microsoft cloud security benchmark (MCSB) is the control set Microsoft Defender for Cloud measures your Azure, AWS and GCP estate against, so it quietly defines what "secure" means in most Azure tenants whether or not anyone has read it. Version 2, published in preview and described by Microsoft as superseding v1, reorganises the benchmark into 12 domains, adds an Artificial Intelligence Security domain, and links controls to more than 420 Azure Policy definitions. This guide explains how the benchmark is structured, what each domain asks for, how it connects to Defender for Cloud and Azure Policy, how to use its mappings to NIST, CIS, ISO 27001, PCI DSS and SOC 2 without overclaiming, and where to supplement it. It is for cloud security architects, GRC teams and auditors. For the general CIS approach see the CIS Controls Guide.
This is an independent guide from NHI Mgmt Group, based on Microsoft's public documentation as of November 2026. It is not sponsored or reviewed by Microsoft. Product features change, so confirm settings against Microsoft's current documentation.
Key takeaways
- v2 is preview. Microsoft's overview labels MCSB v2 as preview and as superseding v1; per-service security baselines for v2 are not yet published, so v1 baselines remain the service-level reference for now.
- Twelve domains, one new. NS, IM, PA, DP, AM, LT, IR, PV, ES, BR, DS and AI; the AI domain has seven controls covering approved models, content filtering, meta-prompts, agent least privilege, human-in-the-loop, monitoring and red teaming.
- It is the default yardstick. Defender for Cloud enables MCSB by default on Azure, AWS and GCP, the classic secure score counts only built-in MCSB recommendations, and exemptions depend on the MCSB initiative being assigned.
- Mappings are a starting point, not evidence. Microsoft states a mapping does not mean you comply with that framework, and some published mappings still cite retired IDs, so verify before reusing them in an audit.
- Identity carries the most weight. IM and PA together hold 16 controls, all but one rated Must have, and IM-3 and IM-8 are the benchmark's non-human identity controls.
What the benchmark is and how it is built
MCSB is Microsoft's prescriptive set of cloud security best practices, successor to the Azure Security Benchmark. Microsoft says it draws on the Cloud Adoption Framework, the Azure Well-Architected Framework, its Secure Future Initiative and CISO Workshop, plus external sources including the AWS Well-Architected Framework, CIS Controls, NIST and PCI DSS. It uses four terms:
- Security domain: a group of related controls, such as Data Protection.
- Security control: a technology-agnostic requirement with an ID, such as DP-1 Discover, classify, and label sensitive data.
- Subcontrol: a numbered implementation step, such as NS-1.1.
- Baseline: control implementations tailored to a service or scenario (for example, an Azure service's security baseline).
Each v2 control page carries a security principle, the risks it mitigates, MITRE ATT&CK mappings, implementation guidance with examples, a link to Azure Policy definitions where one exists, and a criticality level: Must have, Should have or Nice to have. Criticality is the most useful field for planning: it tells you what Microsoft considers baseline versus maturity work.
The 12 domains of MCSB v2
| Prefix | Domain | Scope and representative controls |
|---|---|---|
| NS | Network Security | 10 controls, all Must have: segmentation boundaries (NS-1), network controls for cloud-native services (NS-2), edge firewall, IDS/IPS, DDoS, WAF, central management, insecure protocols (NS-8), private connectivity, DNS security (NS-10) |
| IM | Identity Management | 8 controls, all Must have: centralised identity with isolation (IM-1), protect identity systems (IM-2), application identities (IM-3), service authentication (IM-4), SSO, strong authentication (IM-6), conditional access (IM-7), credential and secret exposure (IM-8) |
| PA | Privileged Access | 8 controls: separate and limit admins (PA-1), no standing access (PA-2), entitlement lifecycle (PA-3, Should have), access reviews, emergency access (PA-5), privileged access workstations (PA-6), just enough administration (PA-7), cloud provider support access (PA-8) |
| DP | Data Protection | 8 controls: discovery and classification, anomaly monitoring, encryption in transit and at rest, customer-managed keys (DP-5, Should have), key management (DP-6), certificate management (DP-7), security of the key and certificate repository (DP-8) |
| AM | Asset Management | Inventory and approval of cloud resources and services |
| LT | Logging and Threat Detection | 7 controls: threat detection (LT-1), identity threat detection (LT-2), investigation logging (LT-3), network logging (LT-4), centralised analysis (LT-5), retention (LT-6, Should have), time synchronisation (LT-7, Should have) |
| IR | Incident Response | Preparation, detection, containment and post-incident activity in cloud |
| PV | Posture and Vulnerability Management | Secure configuration and vulnerability assessment and remediation |
| ES | Endpoint Security | EDR and anti-malware for cloud compute |
| BR | Backup and Recovery | 4 controls: automated backups (BR-1), protect backup data (BR-2), monitor backups (BR-3, Should have), test backups (BR-4, Should have) |
| DS | DevOps Security | 7 controls: threat modelling, software supply chain, DevOps infrastructure (DS-3), SAST, DAST, workload lifecycle, DevOps logging (DS-7, Should have) |
| AI | Artificial Intelligence Security | 7 controls, all Must have: approved models (AI-1), multi-layered content filtering (AI-2), safety meta-prompts (AI-3), least privilege for agent functions (AI-4), human-in-the-loop (AI-5), monitoring and detection (AI-6), continuous AI red teaming (AI-7) |
The AI domain in practice
- AI-1 Ensure use of approved models: a central model registry, automated validation such as hash checks and backdoor scanning, and Azure Policy to deny deployment of models not in the registry.
- AI-2 and AI-3: filter inputs and outputs (Microsoft points to Azure AI Content Safety and Prompt Shields), and use version-controlled system prompts that tell the model to put system instructions above user input.
- AI-5 Ensure human-in-the-loop: define critical actions, such as external data transfers or high-impact decisions, and pause workflows for logged human approval.
- AI-6 and AI-7: detect jailbreaks and prompt injection with Defender for Cloud AI threat protection, correlate in the SIEM using MITRE ATLAS and OWASP references, and red team on a regular cadence with tests embedded in CI/CD.
Control counts and criticality above are taken from the v2 domain pages Microsoft has published; AM, IR, PV and ES are summarised at domain level only. Because v2 is in preview, check the domain page before quoting a control number.
Identity and non-human identity controls
For an identity programme, the IM and PA domains are the core of the benchmark, and they contain practical, testable requirements.
- IM-3 Manage application identities securely and automatically: use managed identities instead of human accounts for applications, choosing system-assigned (lifecycle tied to one resource) or user-assigned (reusable) identities, with least-privilege roles such as Key Vault Secrets User or Storage Blob Data Reader. Where a service principal is unavoidable, prefer certificate credentials over client secrets; if a secret is needed, set an expiry, keep it in Key Vault, rotate before expiry and delete unused ones. See the Azure Managed Identities Security Guide and the Azure Entra Client Secrets Security Guide.
- IM-8 Restrict the exposure of credentials and secrets: secret scanning and removal of embedded credentials, mapped by Microsoft to CIS Safeguards 16.9, 16.10 and 16.12 and NIST IA-5, SC-12, SC-28 among others.
- PA-1 and PA-7: limit Global Administrator and Privileged Role Administrator in Entra ID and Owner, Contributor and User Access Administrator at Azure resource scope; prefer scoped built-in roles.
- PA-2: no standing privilege: Privileged Identity Management eligible assignments with approval, MFA and time-bound activation, and just-in-time VM access through Defender for Cloud.
- PA-5 Set up emergency access: at least two cloud-only break-glass accounts, at least one excluded from Conditional Access, alert on any sign-in, test regularly.
- PA-8: Customer Lockbox so Microsoft support access to data needs customer approval.
- DP-8 and DS-3: protect the vault itself (least-privilege RBAC, PIM, Private Link, soft delete and purge protection) and centralise pipeline secrets with managed identities instead of hard-coded credentials. See the Azure Key Vault Security Guide.
- AI-4 Apply least privilege for agent functions: a capability manifest that denies by default, separate agent identities (Microsoft points to Entra Agent ID) with scoped roles, short-lived scoped tokens and logging of anomalous calls. See the Microsoft Entra Agent ID Guide.
How MCSB connects to Defender for Cloud and Azure Policy
The benchmark is not just a document: in most tenants it is already assigned and scoring.
| Mechanism | What happens | Practical consequence |
|---|---|---|
| Default standard | Enabling Defender for Cloud assigns MCSB on Azure; on AWS MCSB plus AWS Foundational Security Best Practices; on GCP MCSB plus GCP Default | You are being measured against MCSB whether or not you chose it |
| Azure Policy | v2 links controls to more than 420 built-in policy definitions, assigned as an initiative | The initiative is the enforceable part; documentation-only controls are not checked |
| Classic secure score | Only built-in, generally available MCSB recommendations count | Custom and preview checks never move the score |
| Regulatory compliance dashboard | Controls that can be assessed automatically show pass or fail; others show as unavailable | Unavailable controls need manual evidence |
| Exemptions | Defender for Cloud exemptions rely on the MCSB initiative being assigned | Removing the initiative breaks exemption handling and compliance views |
See the Microsoft Defender for Cloud Guide for the plans that unlock the regulatory compliance dashboard, and the Azure Policy and Management Groups Guide for initiative assignment.
Using the framework mappings
Each v2 control lists mappings to NIST SP 800-53 Rev. 5, PCI DSS v4, CIS Controls v8.1, NIST CSF 2.0, ISO/IEC 27001:2022 and the SOC 2 Trust Services Criteria. Used well, these let one set of Azure evidence serve several audits. Used carelessly, they produce findings.
| MCSB control | NIST SP 800-53 Rev. 5 (selection from Microsoft's mapping) | CIS Controls v8.1 (from Microsoft's mapping) | NIST CSF 2.0 (verified ID) |
|---|---|---|---|
| IM-3 Application identities | AC-2, AC-3, IA-2, IA-4, IA-5, IA-9, SC-12, SC-13 | 6.7, 12.5, 16.1, 16.9 | PR.AA-01 |
| IM-8 Credentials and secrets | IA-5, SI-4, RA-5, SC-12, SC-13, SC-28 | 16.9, 16.10, 16.12 | PR.AA-01 |
| PA-2 No standing access | AC-2(1), AC-5, AC-6(2), AC-6(5), AC-16 | 5.4, 6.8 | PR.AA-05 |
| PA-7 Just enough administration | AC-2, AC-3, AC-5, AC-6 | 3.3, 5.4, 6.1, 6.8 | PR.AA-05 |
| DP-3 / DP-4 Encryption in transit and at rest | SC-8, SC-28 | PR.DS-02, PR.DS-01 | |
| LT-3 / LT-5 Logging and central analysis | AU-2, AU-6 | PR.PS-04, DE.AE-03 | |
| BR-2 Protect backup data | CP-9 | PR.DS-11 |
The NIST SP 800-53 and CIS columns for IM-3, IM-8, PA-2 and PA-7 reproduce Microsoft's published mappings; the DP, LT and BR rows and the CSF column are NHI Mgmt Group's own alignment, checked against the NIST catalogues. Three cautions apply:
- Mapping is not compliance: Microsoft says so explicitly. A mapping shows that an Azure feature can help with a requirement, not that the requirement is met.
- Check IDs before reuse: re-map each identifier to CSF 2.0, ISO 27001:2022 Annex A and current ATT&CK before reuse, and confirm with Microsoft where a mapping uses a different identifier style.
- Scope differs: PCI DSS and SOC 2 include people and process requirements that no cloud control set covers.
For cross-framework work see the NIST CSF to ISO 27001 Mapping Guide and the Cloud Security Alliance CCM and STAR Guide.
MCSB compared with other cloud baselines
| Baseline | Nature | Best used for |
|---|---|---|
| Microsoft cloud security benchmark | Control framework with principles, guidance and policy mappings, multi-cloud, built into Defender for Cloud | Day-to-day posture in Azure-centric estates; AI and DevOps coverage |
| CIS Microsoft Azure Foundations Benchmark | Prescriptive, setting-level configuration checks | Configuration audits and third-party assessments |
| CIS Controls v8.1 | Technology-neutral safeguards with implementation groups | Programme prioritisation across on-premises and cloud |
| CSA Cloud Controls Matrix | Cloud control framework for providers and customers | Supplier assurance and STAR |
Choose MCSB as the operating baseline if you live in Defender for Cloud; add the CIS Azure Foundations Benchmark where auditors or customers ask for CIS evidence. See the CIS Cloud Foundations Benchmarks Guide.
Adopting MCSB v2 step by step
- Confirm what is assigned: check which MCSB initiative version is assigned at each management group and which subscriptions and connectors it covers.
- Read the Must have controls: start with IM, PA, DP, LT and NS; assign an owner per domain.
- Map to your obligations: use Microsoft's mappings as a draft, then validate each ID against the current framework text.
- Audit before Deny: run the policy initiative in Audit, review non-compliance by subscription, and agree exemptions with expiry dates.
- Pilot enforcement: move selected policies (for example public network access or TLS minimum) to Deny in a pilot management group, monitor deployment failures, then widen.
- Collect manual evidence: for controls the dashboard shows as unavailable, such as threat modelling (DS-1) or human-in-the-loop (AI-5), define what evidence counts.
- Roll back: revert a Deny assignment to Audit, or add a scoped exemption, if enforcement breaks a critical deployment; record why.
- Track v2 changes: while v2 is preview, review the domain pages periodically and re-check control IDs before audits.
Small versus large organisations
A small team should treat MCSB as a prioritised to-do list: fix the Must have recommendations Defender for Cloud raises on its few subscriptions and ignore the mapping work. A large or regulated organisation should treat it as one input to a unified control framework, keep its own mapping table under version control, and use the initiative only for the automated part.
Measuring progress
- Must have coverage: percentage of Must have controls with an owner and either automated assessment or defined evidence.
- Policy compliance: MCSB initiative compliance per management group, trended monthly.
- Identity controls: standing privileged assignments (PA-2) and service principals with client secrets (IM-3), trending down.
- Exemptions: active exemptions without expiry, target zero.
- Mapping accuracy: mapped requirements re-validated against current framework versions in the last 12 months.
Common failure patterns
- Treating the score as compliance: a high secure score says MCSB recommendations are fixed, not that ISO 27001 or PCI DSS requirements are met.
- Copying mappings into audit evidence: reusing Microsoft's tables without noticing outdated CSF, ISO or ATT&CK identifiers.
- Ignoring documentation-only controls: DS-1, AI-5 and AI-7 have little or no policy coverage, so they never show as failing.
- Blanket exemptions: management group waivers that hide a domain for every future subscription.
- Skipping the AI domain: assuming AI services are covered by general controls when AI-4 asks specifically for agent identities and least-privilege tool access.
- Removing the initiative to quieten noise: this breaks exemptions and the compliance dashboard.
Practitioner checklist
- Record which MCSB version and initiative are assigned at each management group.
- Assign an owner to each of the 12 domains.
- Prioritise Must have controls in IM, PA, DP, LT and NS.
- Replace service principal secrets with managed identities or certificates under IM-3.
- Remove standing privileged access under PA-2 and test break-glass accounts under PA-5.
- Re-validate every reused mapping against NIST CSF 2.0, ISO 27001:2022 and ATT&CK v20.
- Define manual evidence for controls the dashboard cannot assess.
- Put an expiry and ticket on every exemption.
- Review the AI domain for every AI and agent workload.
Standards and references
- Microsoft Learn: Microsoft cloud security benchmark v2 overview
- Microsoft Learn: Introduction to the Microsoft cloud security benchmark
- Microsoft Learn: MCSB v2 Identity Management
- Microsoft Learn: MCSB v2 Privileged Access
- Microsoft Learn: MCSB v2 Artificial Intelligence Security
- Microsoft Learn: Regulatory compliance standards in Defender for Cloud
- NIST: Cybersecurity Framework 2.0
- NIST: SP 800-53 Rev. 5
Related NHI Mgmt Group resources: CIS Controls Guide · Microsoft Defender for Cloud Guide · Microsoft Azure Security Guide · Azure Landing Zones Security Guide