Most Microsoft 365 tenants have years of unlabelled, overshared content, and an AI assistant that can read everything a user can read turns that debt into an immediate exposure. Microsoft Purview Information Protection is the sensitivity label, auto labelling and encryption layer that lets an organisation mark what is sensitive, enforce access through Rights Management, and feed data loss prevention (DLP) policies and Copilot with those signals. This guide covers how labels decide access, which Purview and Entra roles control the system, the service accounts and super users that can decrypt everything, what to log, and a hardening baseline with a phased rollout. It is written for Microsoft 365 administrators, data security engineers and security architects preparing a tenant for Microsoft 365 Copilot. It describes product features from Microsoft documentation and does not recommend any vendor.
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
- Labels are metadata plus optional encryption. A sensitivity label is stored as clear-text metadata in the file or email; only labels configured with encryption change who can open the content, and Copilot respects that encryption only through usage rights such as EXTRACT.
- Auto labelling is two different engines. Client-side labelling runs in Office apps as users work and can recommend; service-side policies label SharePoint, OneDrive and Exchange at scale, must run in simulation first, and never overwrite manual labels unless you configure it.
- The super user feature is the master key. Rights Management super users get Full Control over every encrypted item, Global Administrators can grant it to themselves, and decryptions by super users are not flagged as such in usage logs.
- Entra roles silently override Purview scoping. A user with an Entra role and an administrative-unit scoped Purview role group gets unscoped access, so role design needs PIM for Groups and no stacking.
- Labelling comes before Copilot, not after. Label the most sensitive repositories, fix container sharing settings and add DLP conditions on labels before broad Copilot licensing, because the assistant surfaces whatever permissions already allow.
How identity and access work in Purview Information Protection
Purview Information Protection layers three controls on top of normal Microsoft 365 permissions. Understanding which one actually stops access is the most common gap in deployments. For the classification scheme itself, see the Data Classification Guide; this page covers the Microsoft implementation.
| Control | What it governs | Identity it evaluates | Limits to know |
|---|---|---|---|
| Item label without encryption | Visual marking (header, footer, watermark), DLP and Copilot signals | None: anyone with file access can open it | Watermarks up to 255 characters, headers and footers up to 1,024 (255 in Excel); one sensitivity label per item |
| Item label with encryption | Who can open, edit, print, copy or extract, and for how long, enforced by Azure Rights Management | Entra ID users and groups, or "any authenticated user", with usage rights | Labels with specified users and permissions are capped at 500 per tenant; encryption follows the file outside the tenant |
| Container label | Privacy, external sharing, guest access and unmanaged device access for Teams, Microsoft 365 Groups, SharePoint sites, Viva Engage and Loop workspaces | Site and group membership, device state through Conditional Access | Labels the container, not the items in it; must be enabled before use |
- Label scopes: each label applies to Files and other data assets, Emails, Meetings (requires both of the former) or Groups and sites. A label with the wrong scope silently fails to appear or to auto-apply.
- Priority: label order is priority, with the most restrictive label at the bottom. Priority drives downgrade justification and which label wins when several auto labelling rules match.
- Label policies: labels reach users through publishing policies targeted at users and groups. A policy can set a default label for documents, email, new containers and Power BI content, require justification for downgrades and make labelling mandatory. Where a user is in several policies, the policy with the highest order number wins conflicting settings. Changes can take up to 24 hours to replicate.
- Taxonomy size: Microsoft recommends no more than five main labels with five sublabels each. Parent labels are being replaced by label groups, and parent labels must never be the default or auto-applied label.
- Copilot behaviour: Microsoft 365 Copilot and Copilot Chat recognise labels, show the highest-priority label among the items used in a response, and return data from an encrypted item only if the user holds the EXTRACT usage right. A label without encryption does not stop Copilot from reading content the user can already open. The data at risk is whatever the assistant can retrieve for the user, which is why scope reduction comes first. See the Microsoft 365 Copilot Security Deep Dive.
Choosing between label approaches
| Approach | Choose when | Trade-off |
|---|---|---|
| Default label plus manual labelling | Small tenants, or the first months of a rollout | Cheap and fast, but accuracy depends on users and training |
| Client-side auto or recommended labels | Content created daily in Word, Excel, PowerPoint and Outlook | Applies before save with minimal delay; needs supported app versions; cannot label existing data |
| Service-side auto labelling policies | Labelling data at rest in SharePoint and OneDrive or mail in transit | Scales (500,000 files per tenant per day), requires E5-tier licensing and simulation, no user prompt |
| Encrypting labels | Regulated or highly confidential content that may leave the tenant | Strongest control, but breaks some search, indexing and third-party tools; avoid as a default document label |
| Container labels only | Oversharing is the main risk and site owners are reliable | Fixes sharing settings at the site level without touching items |
Admin roles and privileged access
Configuring labels and policies needs Purview permissions; applying labels in apps does not. The privileged question is who can change a label's encryption, see matched content or decrypt items.
| Role or role group | What it allows | Recommended handling |
|---|---|---|
| Information Protection Admins (Sensitivity Label Administrator role) | Create and edit labels and label policies | Small group, PIM for Groups activation, change approval for encryption settings |
| Information Protection Analysts, Investigators, Readers | Activity explorer, content review, read-only reporting | Investigators see content: treat as privileged |
| Data Classification Content Viewer | View the actual matched content in auto labelling simulation and content explorer | Not held by Global Administrators by default; grant time-bound |
| Compliance Administrator, Compliance Data Administrator | Turn auto labelling policies on, manage DLP | Use only where the narrower IP roles do not cover the task |
| Global Administrator, Azure RMS GlobalAdministrator | Can enable the super user feature and add themselves | Minimise holders, alert on super user changes |
| Rights Management super user | Full Control over all content encrypted by the tenant, including content protected before assignment | Feature disabled except during a documented task |
- Scoping trap: Purview role groups can be scoped to administrative units, but if the same user also holds an Entra role such as Global Reader, the Entra role wins and access becomes unscoped. Do not stack the two on one person.
- Temporary assignments: Purview supports expiry on role group assignments from 1 day to 2 years, and PIM for Groups lets Entra security groups be assigned to role groups for just-in-time activation. Direct user assignments cannot use just-in-time activation, so assign through groups. See the Privileged Identity Management (PIM) Guide.
Integrations, API tokens and service accounts
Information protection depends on several non-human identities that can read or decrypt large volumes of content. They belong in the NHI inventory with an owner and a review date.
- Information protection scanner service account: the scanner that labels on-premises file shares and SharePoint Server runs under an Active Directory account synchronised to Entra ID. It needs Log on as a service; Log on locally is required only for install and configuration and can be removed afterwards. On file shares it needs Read, Write and Modify (Read only in discovery mode), on SharePoint Full Control, and it must be a super user if labels reprotect or remove encryption. It also holds db_owner on the AIPScannerUL_<cluster> database.
- Unattended labelling: where the service account cannot sync to Entra ID, Set-Authentication with the DelegatedUser and OnBehalfOf parameters runs labelling cmdlets under a separate Entra account. That Entra account is a credential to vault and monitor.
- Services assigned as super users: DLP, content encryption gateways, anti-malware and archiving tools may be added with Add-AipServiceSuperUser so they can inspect encrypted files. Each one can decrypt everything.
- Applications using the MIP SDK: third-party apps read clear-text label metadata and can apply labels through the Microsoft Information Protection SDK, typically under an Entra app registration. Review their Graph and Rights Management permissions like any other high-privilege app. See the Entra ID Applications and Consent Guide.
- Copilot agents: an agent configured to run with its author's or a service identity's permissions, rather than the end user's, can surface labelled content to people who could not open it directly. See the Securing Microsoft Copilot Studio Agents Guide.
Logging and detection
Labelling and DLP events land in the Microsoft 365 unified audit log and are presented in Activity explorer, which holds 30 days and lags 60 to 90 minutes for core services. Export to a SIEM for anything longer. See the Microsoft 365 Unified Audit Log Guide.
| Signal | Source | What to detect |
|---|---|---|
| Label applied, Label changed (downgraded or removed) | Activity explorer and audit log | Bulk downgrades by one user, removal of encrypting labels before download or sharing |
| Protection changed, Protection removed | Information protection client and scanner | Encryption stripped outside a change window |
| File read on labelled content | Activity explorer | Unusual volume of Highly Confidential reads, including by service identities |
| DLP rule matched, DLP rule undo | Audit log, DLP alerts (30 days in Purview, six months in the Defender portal) | Repeated overrides with weak justifications, matches in Copilot interactions |
| Super user feature and membership changes | Get-AipServiceAdminLog | Enable-AipServiceSuperUserFeature or Add-AipServiceSuperUser outside a ticket |
| Super user decryption | Rights Management usage logs cross-referenced with Get-AipServiceSuperUser | Any decryption by a super user, since logs do not mark it |
| Label and policy configuration changes | Audit log | Encryption settings removed from a label, mandatory labelling turned off |
These map to ATT&CK collection and exfiltration techniques such as T1213.002 Sharepoint, T1114 Email Collection and T1567 Exfiltration Over Web Service. For the DLP design side, see the Data Loss Prevention (DLP) Guide.
Metrics:
- Percentage of files in priority SharePoint sites carrying a label, and percentage of Highly Confidential content that is encrypted.
- Label downgrade rate per 1,000 users per month, and the share with a meaningful justification.
- Number of sites and teams without a container label, and number with "anyone" sharing links.
- DLP override rate for label-based rules.
- Number of super users and super user service identities, with days the feature was enabled (target: zero standing).
- Auto labelling false positive rate from simulation review, before and after rule tuning.
Hardening baseline
- Prerequisites: confirm unified auditing is on, enable sensitivity labels for Office files in SharePoint and OneDrive (otherwise the service cannot read encrypted files for search, co-authoring or auto labelling), enable PDF support if needed, and confirm licensing (service-side auto labelling needs an E5 or equivalent tier).
- Taxonomy: publish four or five labels (for example Public, General, Confidential, Highly Confidential) with tooltips tested with users. Put the most restrictive label last. Use label groups rather than parent labels.
- Pilot ring: publish labels to IT and a pilot business unit with a default label of General (non-encrypting) and justification on downgrade. Measure accuracy for four to six weeks.
- Containers: enable container labels and attach external sharing, guest and unmanaged-device settings to Confidential and Highly Confidential sites before Copilot licensing.
- Auto labelling in simulation: start with one SharePoint site, then more sites, OneDrive, then Exchange. Run simulation (up to 12 hours), review with a time-bound Data Classification Content Viewer, refine sensitive information types, and repeat. Policies that match more than 20,000,000 files cannot be turned on: narrow scope. Encrypting labels used for SharePoint auto labelling must use "Assign permissions now" with no expiry.
- Enforce: turn on policies, decide deliberately whether to overwrite lower-priority manual labels, and add DLP rules that use "content contains sensitivity label" for external sharing, endpoint copy to USB and the Microsoft 365 Copilot location (preview at the time of writing). Run DLP in simulation mode first.
- Privilege: move label administration to PIM for Groups, keep the super user feature disabled, and vault the scanner and unattended labelling credentials.
- Rollback: deleting or editing a label does not remove it or its protection from existing content, so rollback means publishing a corrected label, using a removal policy or bulk relabelling with Set-FileLabel under a temporary super user. Test that path before enforcing encryption at scale.
Small versus large tenants: a small tenant on Business Premium can rely on manual labelling, a default label, container labels and DLP; service-side auto labelling is not included in that licence. Large tenants need service-side policies, administrative-unit scoping and a scanner cluster for file shares.
Framework mapping
| Framework | Reference | Purview capability |
|---|---|---|
| NIST SP 800-53 Rev. 5 | AC-16 Security and Privacy Attributes | Sensitivity labels as attributes bound to content |
| NIST SP 800-53 Rev. 5 | AC-3 Access Enforcement; SC-28 Protection of Information at Rest | Encrypting labels and Rights Management |
| NIST SP 800-53 Rev. 5 | AC-6(5) Privileged Accounts; AU-6 Audit Record Review | Role design, super user control, label activity review |
| NIST SP 800-53 Rev. 5 | SC-7(10) Prevent Exfiltration | DLP rules using labels |
| NIST CSF 2.0 | ID.AM-05; PR.DS-01; PR.DS-02; PR.AA-05; DE.CM-03 | Classification, protection at rest and in transit, least privilege, monitoring |
| ISO/IEC 27001:2022 Annex A | 5.12 Classification of information; 5.13 Labelling of information; 8.12 Data leakage prevention | Taxonomy, labels, DLP |
| CIS Controls v8.1 | Safeguard 3.7 Establish and Maintain a Data Classification Scheme | Label taxonomy and policy |
Common misconfigurations
- Encrypting default label for documents. Microsoft warns against it because external recipients may not be able to open files. Default to a non-encrypting label and encrypt by choice or auto labelling.
- Labels without encryption treated as access control. A Highly Confidential marking on a file shared to "everyone except external users" stops nothing, including Copilot. Fix permissions and container settings, then encrypt what truly needs it.
- Standing super users. Leaving the feature enabled with archiving or DLP service identities as permanent members creates an identity that can decrypt every item without a distinctive log record.
- Turning on auto labelling without simulation review. Simulation is required, but teams skip the content review, ship a noisy sensitive information type and label thousands of files wrongly. Changing a custom type later does not reclassify existing files.
- Scanner account with Log on locally and Domain Admin rights. The scanner needs repository permissions, not domain-wide privilege. Remove Log on locally after setup.
- Exchange auto labelling scoped to selected senders. Narrowing the sender scope exempts external senders, so incoming sensitive mail goes unlabelled. Keep All included and filter with advanced rules.
Practitioner checklist
- Confirm auditing, SharePoint and OneDrive label support, PDF support and licensing before publishing any label.
- Publish five or fewer main labels, most restrictive last, with a non-encrypting default.
- Apply container labels with sharing and unmanaged-device settings to sensitive sites before Copilot rollout.
- Run every service-side auto labelling policy in simulation, review content with a time-bound viewer role and tune.
- Add DLP rules that use sensitivity labels as conditions, including the Copilot location where licensed.
- Assign Information Protection roles through PIM for Groups and avoid stacking Entra roles on scoped users.
- Keep the Rights Management super user feature disabled and alert on Enable-AipServiceSuperUserFeature.
- Inventory the scanner account, unattended labelling account and super user services as NHIs with owners.
- Export label and DLP events to the SIEM, since Activity explorer holds 30 days.
- Test the rollback path (relabel and decrypt) before enforcing encryption at scale.
Standards and references
- Microsoft Learn: Learn about sensitivity labels
- Microsoft Learn: Apply a sensitivity label to content automatically
- Microsoft Learn: Configuring super users for Microsoft Purview Information Protection
- Microsoft Learn: Permissions in the Microsoft Purview portal
- Microsoft Learn: Requirements for the information protection scanner
- Microsoft Learn: Learn about data loss prevention
- Microsoft Learn: Get started with activity explorer
- NIST: SP 800-53 Rev. 5, Security and Privacy Controls
Related NHI Mgmt Group resources: Data Classification Guide · Microsoft 365 Copilot Security Deep Dive · Data Loss Prevention (DLP) Guide · SharePoint and OneDrive Security Guide · Microsoft 365 Security Guide