An administrator who reads email and browses the web on the same laptop they use to manage the domain, the identity provider or the cloud tenant has given every phishing link a path to the crown jewels. A privileged access workstation (PAW) is a dedicated, locked-down device used only for privileged tasks, so that the device an admin session starts from is at least as trustworthy as the system being administered. This guide covers the clean source principle, PAW design options, a concrete build baseline, cloud-hosted PAWs, rollout and rollback, detection and the mistakes that undo the investment. It is written for infrastructure, identity and endpoint engineers building PAWs, and for security architects deciding which roles need one. It goes deeper than the Tiered Administration and Enterprise Access Model Guide, which explains the tier model a PAW programme sits inside.
Key takeaways
- The device is a security dependency. Whoever controls the device controls every session started from it, so MFA, PAM vaults and session recording cannot compensate for a compromised admin endpoint.
- Browse down, never up. Administer a system only from a device trusted at least as much as that system; an everyday productivity laptop is never a valid source for Tier 0 or cloud control plane work.
- Deny by default is the defining feature. A PAW blocks general web browsing and email, runs only approved applications, removes local admin rights and allows outbound traffic only to the management endpoints the role needs.
- Enforce it at the target, not just on the device. Conditional Access device filters, authentication policies and firewall rules must refuse privileged sign-ins from anything that is not a PAW, or admins will drift back to their laptops.
- Scope it to the roles that matter. Start with identity provider, directory, hypervisor, backup and cloud tenant administrators, then extend; a PAW for every IT user too early creates workarounds.
Why the admin device matters: the clean source principle
Microsoft's privileged access guidance states the clean source principle as a requirement that all security dependencies be as trustworthy as the object being secured. Any subject that controls an object is a security dependency of it. An admin's workstation controls every keystroke, token and session that passes through it, so it is a dependency of every system that admin manages. If an attacker owns the workstation, they can impersonate the admin or steal credentials for later use, whatever controls sit downstream.
The UK NCSC's secure system administration guidance makes the same point with different words. Its first principle is to gain trust in your management devices. It calls administering a system from a less trusted device "browse up" and treats it as an anti-pattern; "browse down", administering only from a device trusted at least as much as the target, is the recommended model. A device that receives email is less trusted because it is exposed to phishing; a device with no direct internet access is more trusted.
Three attack paths explain why this is not theoretical:
- Credential theft from memory: once an attacker has local admin on a workstation, credentials and tickets of anyone who signs in are exposed (ATT&CK T1003.001 LSASS Memory), then replayed elsewhere (T1550.002 Pass the Hash, T1550.003 Pass the Ticket). See the Pass-the-Hash and Pass-the-Ticket Guide.
- Session and token theft: browser cookies and refresh tokens for admin portals can be lifted from a compromised browser profile and reused without the password or MFA.
- Stored credentials: passwords saved in a browser or a synced profile follow the user to any device. The Okta support system breach began with a support service account whose credentials had been saved to an employee's personal Google profile in Chrome on a work laptop, the opposite of a clean source.
Microsoft's intermediaries guidance is blunt that PAM tooling "does not address many privileged access risks, notably risk of device compromise". A vault that injects a password into a session started from an infected laptop still hands that session to the attacker.
Who needs a PAW
Microsoft's model defines three device security levels: enterprise (any application, any website, EDR required), specialised (no self-administration, only admin-installed applications) and privileged (the PAW, for roles where account compromise would have a significant or material impact). Map roles to levels by what an account can change, not by job title.
| Role or access | Device level | Reason |
|---|---|---|
| Domain and enterprise admins, AD CS and AD FS administrators | Privileged (PAW) | Control of Tier 0 means control of every identity |
| Identity provider super admins (Entra Global Administrator, Privileged Role Administrator, Okta Super Admin) | Privileged (PAW) | Can create users, reset MFA and change federation |
| Cloud tenant owners, organisation admins, root and break-glass accounts | Privileged (PAW) | Control plane for all workloads and workload identities |
| Hypervisor (ESXi, vCenter), backup and storage administrators | Privileged (PAW) | Ransomware operators target these to destroy recovery |
| Secrets vault, PAM and CI/CD platform administrators | Privileged (PAW) | Hold credentials for most other privileged identities |
| Server and application administrators (Tier 1) | Specialised or PAW | High impact on business systems; PAW if they reach Tier 0 paths |
| Help desk, workstation support (Tier 2) | Specialised | Broad reach but bounded impact; still no browsing on admin sessions |
| Developers and data engineers with production access | Specialised, PAW for production secrets | See the Developer Workstation and IDE Security Guide |
Small organisations: with two or three administrators, a pair of dedicated, cheaply specified laptops used only for the identity provider, the cloud console and backups is a realistic first step. NCSC's guidance accepts that a small organisation may tolerate browse-up risk where compromise impact is low, but that judgement rarely holds for the identity provider admin.
Large organisations: run separate PAW builds per tier, a dedicated device group and management scope, and treat the PAW estate as a Tier 0 asset in its own right.
PAW design options compared
| Model | How it works | Strengths | Weaknesses | Fits |
|---|---|---|---|---|
| Dedicated physical PAW plus separate productivity device | Two devices; the PAW has no email or general web | Strongest separation; simple to reason about | Cost, desk space, admin frustration, two devices to carry | Tier 0 and cloud control plane admins |
| PAW host with productivity in a virtual machine or cloud PC | The hardened PAW is the host; the less trusted workload runs as a guest or remote desktop | One device; host stays clean because email runs elsewhere | Guest escape risk is low but non-zero; more engineering | Admins who need both all day |
| Productivity host with admin VM | Admin VM runs on the everyday laptop | Cheap | Browse up: the less trusted host controls the admin VM. Not a PAW | Not recommended for privileged roles |
| Cloud-hosted PAW (cloud PC or VDI) reached from a PAW-grade endpoint | Admin desktop runs in the cloud, locked down like a PAW | Fast provisioning, central control, no local data | The endpoint that connects to it is still a dependency; it must be specialised or PAW grade | Distributed teams, contractors with strict controls |
| Jump server or bastion only | Admins RDP or SSH from laptops to a shared hardened host | Central logging, network choke point | Credentials and sessions pass through a shared host; the starting laptop is unchanged | An intermediary, not a replacement for PAWs |
The rule for cloud PAWs: a cloud PC or virtual desktop moves the admin desktop off the laptop, but the device that displays it and captures keystrokes is still a dependency. Treat the connecting device as at least specialised, enforce compliant-device Conditional Access on the cloud PC sign-in, and apply the same deny-by-default build inside the cloud desktop. Microsoft notes that the Azure portal used to reach Azure Bastion must meet the security level of the resources reached through it, typically privileged or specialised.
Jump servers: Microsoft's intermediaries guidance describes jump servers as attractive targets because many sensitive sessions pass through them and they are often neglected and internet exposed. Use them as a controlled network path from PAWs to legacy targets, never as the reason to skip PAWs.
Build baseline for a PAW
The baseline below draws on Microsoft's privileged device profile and deployment guidance, which is written for Windows managed with Intune and Entra ID. The same principles apply to macOS or Linux PAWs with the equivalent controls.
Hardware and boot integrity
- New hardware from the OEM: start from a new device built to your specification and provisioned with Windows Autopilot, so no prior image or user has touched it.
- Root of trust: TPM 2.0, UEFI Secure Boot, virtualisation and HVCI (memory integrity) enabled, HVCI-ready drivers, DMA I/O protection and System Guard, with drivers and firmware from Windows Update.
- Disk encryption: BitLocker on all local disks with TPM protection and recovery keys escrowed to the directory.
- Device health: use Device Health Attestation and compliance policy so Conditional Access only accepts devices that booted cleanly.
Credential and OS protection
- Credential Guard and LSA protection: virtualisation-based security isolates LSASS secrets, which blunts T1003.001 even if malware runs.
- No local administrators: set "Additional local administrators on Microsoft Entra joined devices" to None, use a standard user account type in the Autopilot profile, and manage the built-in admin with Windows LAPS. See the Local Administrator Password Solution (LAPS) Guide.
- Application control: Microsoft's privileged profile enforces AppLocker; App Control for Business (formerly WDAC) gives stronger, kernel-enforced control. Allow only the admin tools the role needs, installed by the management platform. Run new rules in audit mode first.
- Attack surface reduction: remove legacy components (Microsoft's deployment script removes PowerShell 2.0 and Internet Explorer), prevent saving network credentials and apply a security baseline.
Network and web restrictions
- Firewall: Microsoft's privileged profile blocks inbound and outbound traffic by default and does not merge local rules; outbound is allowed only for HTTP/HTTPS (80/443), DNS (53), DHCP and Windows Time (123).
- Allow-listed destinations: the privileged profile sets URL filtering to deny by default. Microsoft's example uses a local proxy setting with a proxy-override list of management endpoints so all other web traffic fails. A secure web gateway with a per-role allow list is the more maintainable equivalent.
- No email, no general browsing, no personal profiles: no mail client, no browser sync to personal accounts, no consumer cloud storage.
Enrolment and management
- Restrict who can join: allow only the PAW users group to join devices, require MFA to join, block personally owned enrolment and use an Enrolment Status Page that blocks use until all policies apply.
- Isolated management scope: put PAWs in their own dynamic device group (Microsoft's example tags devices with an order ID such as "[OrderID]:PAW") and scope Intune admin roles so that ordinary endpoint admins cannot push software to PAWs. The management platform is a security dependency of every PAW. The Stryker Intune wiper attack shows what one hijacked device management admin account can do to an estate.
- Patching: fast update rings with short deferrals; PAWs should be among the first devices patched, not the last.
Enforcing PAW use at the target
A PAW that admins can bypass becomes an expensive spare laptop. Enforcement belongs on the resources being administered.
| Target | Enforcement control | What it blocks |
|---|---|---|
| Entra ID and Microsoft cloud admin portals | Conditional Access with device filters or compliant-device requirement for privileged roles; phishing-resistant authentication strength | Admin sign-in from non-PAW devices, token replay from other devices |
| Other identity providers and SaaS admin consoles | Device trust or managed-device policies scoped to admin roles; admin console network zones | Super admin access from laptops and personal devices |
| Active Directory Tier 0 | Authentication policy silos and "Deny log on" user rights for Tier 0 accounts on non-Tier 0 hosts; firewall rules allowing RDP, WinRM and RSAT ports only from PAW subnets | Domain admin logons on servers and workstations, lateral movement to DCs |
| Hypervisors, backup consoles, network devices | Management network reachable only from PAW or jump host subnets | Direct access from user VLANs |
| Cloud provider consoles and CLIs | Source IP or device conditions on admin roles; break-glass excluded and monitored | Console and CLI use from unmanaged devices |
Exclude emergency access accounts from device-based policies, as Microsoft's deployment guide does with its Emergency BreakGlass group, and alert on every use. See the Break-Glass Account Guide and the Conditional Access Guide.
Where identity fits
A PAW only works when the accounts used on it are separate as well. Each administrator gets a dedicated admin account (CIS Controls v8.1 Safeguard 5.4) used only on the PAW, with phishing-resistant MFA such as FIDO2 security keys or certificate-based authentication, as CISA's Scattered Spider advisory (AA23-320A) recommends for accounts with access to critical systems. Privilege should be activated just in time on the PAW rather than standing; see the JIT Access Guide.
Non-human identities need the same thinking. Service account passwords, cloud access keys and API tokens used for admin automation should be retrieved from a vault onto the PAW or, better, used only by pipelines running on hardened build agents. A long-lived cloud key copied into a script on a productivity laptop defeats the whole design. See the Service Account Security Guide.
Implementation roadmap
- Scope (weeks 1 to 2): list Tier 0 and control plane roles and the accounts that hold them, including vendor and break-glass accounts. Inventory the tools each role actually uses and the endpoints those tools contact.
- Build the management plane (weeks 2 to 4): create the PAW device group, a dedicated management scope and a small set of PAW administrators who themselves use PAWs.
- Build and audit (weeks 4 to 6): deploy the baseline to five to ten pilot devices with application control and web restrictions in audit mode. Collect blocked application and URL events.
- Pilot ring (weeks 6 to 10): issue PAWs to identity and cloud admins. Turn application control and web allow lists to enforce one role at a time. Fix gaps within a fixed service level so admins do not abandon the device.
- Enforce at targets (weeks 8 to 12): switch Conditional Access device filters and AD logon restrictions from report-only to enforced for the pilot roles. Keep break-glass excluded and monitored.
- Expand (months 3 to 9): extend to hypervisor, backup, vault and CI/CD admins, then Tier 1 server admins on specialised or PAW builds.
- Operate: quarterly review of allowed applications and URLs, hardware refresh into the build, and annual red team testing of the PAW boundary.
Rollback: keep each enforcement step separately reversible. Application control policies can return to audit mode; Conditional Access policies can return to report-only; logon restrictions are group based, so removing an account from the restricted group restores access. Document a time-limited exception process run through the change board, rather than letting admins revert to laptops informally.
Detection and measurement
Monitor both the PAW estate and the targets for evidence of bypass.
- Privileged sign-ins from non-PAW devices: identity provider sign-in logs filtered to admin roles where the device is not in the PAW group or not compliant. In Active Directory, Event ID 4624 for Tier 0 accounts on hosts outside the PAW and DC set, and 4672 (special privileges assigned) on unexpected hosts.
- Application control events: AppLocker and App Control block and audit events (Microsoft's guide queries Defender Advanced Hunting with
DeviceEvents | where ActionType startswith "AppControl"). - Policy drift: changes to PAW Intune profiles, compliance policies, the PAW device group or Conditional Access exclusions should page the security team.
- Remote access tooling: any remote support or RMM agent on a PAW is an incident (ATT&CK T1219 Remote Access Tools).
| Metric | Target | Why it matters |
|---|---|---|
| Share of Tier 0 and control plane admin sign-ins from PAWs | 100% except logged break-glass | The core outcome |
| Privileged accounts in scope with an issued PAW | 100% of in-scope roles | Coverage |
| PAWs compliant with baseline (encryption, Credential Guard, app control enforced) | Above 98% | Build integrity |
| Median time to approve a new admin tool or URL | Under 5 working days | Usability; slow approvals create workarounds |
| Active PAW exceptions older than 30 days | Zero | Exceptions become permanent |
Framework mapping
| Framework | Reference | PAW relevance |
|---|---|---|
| NIST SP 800-53 Rev 5 | AC-6(2) Non-privileged Access for Nonsecurity Functions; AC-6(5) Privileged Accounts; AC-17 Remote Access | Separate admin accounts and controlled admin paths |
| NIST SP 800-53 Rev 5 | CM-6 Configuration Settings; CM-7 Least Functionality; SC-7 Boundary Protection | Hardened build, application control, egress allow list |
| NIST SP 800-53 Rev 5 | AC-6(9) Log Use of Privileged Functions; SI-4 System Monitoring | Detection of bypass and drift |
| NIST CSF 2.0 | PR.AA-05; PR.PS-01; DE.CM-09 | Access policy, configuration management, monitoring |
| CIS Controls v8.1 | 5.4 Restrict Administrator Privileges to Dedicated Administrator Accounts; 12.8 Establish and Maintain Dedicated Computing Resources for All Administrative Work | 12.8 is the PAW control in CIS terms |
| ISO/IEC 27001:2022 Annex A | A.8.1 User endpoint devices; A.8.2 Privileged access rights | Device and privileged access policy |
| MITRE ATT&CK v20 | T1003.001, T1550.002, T1550.003, T1566, T1078.002 | Techniques a PAW is designed to break |
Common failure patterns
- The PAW in a drawer: devices are issued but nothing at the target requires them, so admins keep using laptops. Enforcement at the identity provider and in AD is what makes the programme real.
- Browse-up virtualisation: an "admin VM" on the productivity laptop is presented as a PAW. The host controls the guest, so this is the anti-pattern the NCSC warns against.
- Shared management plane: PAWs are managed by the same Intune or SCCM admins and software distribution as every other device, so a compromise of ordinary endpoint management reaches Tier 0.
- Web allow list erosion: exceptions accumulate until the PAW can reach webmail and file sharing sites. Review the list quarterly and tie every entry to a role.
- Help desk bypass: attackers call the service desk, reset an admin's MFA and enrol their own device. In the MGM Resorts breach the attackers claimed a help desk call led to Okta and Azure admin rights. PAW device enforcement plus strict verification for admin resets closes that path; see the Account Recovery and Help Desk Security Guide.
- Credentials outside the PAW: admin passwords saved in personal browser profiles or password managers on everyday devices, as in the Okta case above.
Practitioner checklist
- List every Tier 0, identity provider, cloud tenant, hypervisor, backup and vault admin role and map it to a device level.
- Provision PAWs from new hardware with Autopilot, TPM 2.0, Secure Boot, HVCI, Credential Guard and BitLocker.
- Remove local admin rights and enforce application control after an audit-mode period.
- Block email and general web browsing; allow outbound traffic only to role-specific management endpoints.
- Place PAWs in a dedicated device group and management scope administered from PAWs.
- Issue separate admin accounts with phishing-resistant MFA, used only on PAWs.
- Enforce PAW-only access at the identity provider, AD and management networks, with break-glass excluded and alerted.
- Alert on privileged sign-ins from non-PAW devices and on changes to PAW policies.
- Track the five metrics above monthly and close exceptions within 30 days.
Standards and references
- Microsoft: Securing devices as part of the privileged access story
- Microsoft: Privileged access deployment
- Microsoft: Success criteria for privileged access strategy
- Microsoft: Securing privileged access intermediaries
- NCSC: Secure system administration
- NCSC: Gain trust in your management devices
- CISA: Scattered Spider advisory AA23-320A
Related NHI Mgmt Group resources: Tiered Administration and Enterprise Access Model Guide · Privileged Access Management Guide · Windows Endpoint Hardening Guide · Privileged Session Management Guide · Privileged Account Discovery and Onboarding Guide