An enterprise certificate authority is an identity issuer for every machine, workload and smart card that trusts it, so a badly designed hierarchy turns one stolen key into a forged identity for anything in the estate. Most internal PKIs grew by accident: a single online root installed years ago, templates copied between projects, CRL locations that nobody owns. This guide covers how to design the hierarchy itself: offline roots, policy and issuing CAs, how to constrain each CA, key protection, revocation publishing, and the certificate policy (CP) and certification practice statement (CPS) that make the design auditable. It is for PKI architects, identity engineers and security teams responsible for device, workload and user certificates. Certificate lifecycle and automation are covered separately in the Machine Identity, PKI and Certificate Lifecycle Guide, and basics in the PKI Basics Guide.
Key takeaways
- Keep the root offline and boring. The root signs only subordinate CA certificates and CRLs, lives on an HSM or an air-gapped machine, and comes out a few times a year under a scripted ceremony.
- Separate issuing CAs by purpose and risk. Domain authentication, Wi-Fi and VPN devices, TLS servers, code signing and workload identity should not share one issuing key.
- Constrain every subordinate. Use basic constraints with a path length, extended key usage, name constraints and certificate policy OIDs so a compromised CA can only issue what it was meant to.
- Revocation infrastructure is part of the design. AIA and CDP locations are baked into every certificate at issuance, so they must be decided, reachable over HTTP and owned before the first certificate is issued.
- Write the CP and CPS. RFC 3647 gives the nine-part outline; without it the PKI cannot be audited, handed over or trusted by partners.
Hierarchy models
| Model | Structure | Strengths | Weaknesses | Fits |
|---|---|---|---|---|
| One tier | Online root issues end-entity certificates | Simple, cheap | Root key exposed on a network server; compromise means rebuilding all trust | Labs and short-lived test environments only |
| Two tier | Offline root, one or more online issuing CAs | Root protected; issuing CAs can be revoked and replaced independently | Policy separation relies on templates and constraints at the issuing tier | Most organisations |
| Three tier | Offline root, offline policy (intermediate) CAs, online issuing CAs | Clear policy or regional separation; a policy CA can be revoked without touching others | More ceremonies, longer chains, more CRLs to keep current | Large, multinational or regulated estates; partner cross-certification |
| Cloud-managed private CA | Provider-hosted root and subordinates with HSM-backed keys | No hardware to run; API issuance for workloads | Provider becomes part of your trust base; check key export, logging and region | Cloud-native workloads, service mesh; see the AWS Certificate Manager and Private CA Guide |
A two-tier hierarchy is the default recommendation in most enterprise designs. Move to three tiers when you need separately revocable branches, for example different assurance levels, legal entities or regions, or when an external party will cross-certify one branch only.
Designing the root CA
- Offline means powered off and unnetworked: a standalone (non-domain-joined) root that never touches the production network. Requests and signed certificates move by removable media under dual control.
- Key protection: generate the root key inside an HSM validated to FIPS 140-3 where possible. NIST's Cryptographic Module Validation Program states that FIPS 140-2 validated modules could be accepted for new systems only until 21 September 2026; they are now on the Historical list and usable only in existing systems. Microsoft's AD CS guidance is to install and configure the HSM before creating any CA whose keys it will hold.
- Algorithm and size: choose the strongest options all relying parties support. RSA 4096 or ECDSA P-384 with SHA-256 or SHA-384 are common root choices; avoid SHA-1 and MD5, which AD CS still lists for legacy providers. Plan for post-quantum migration by keeping the hierarchy easy to re-root.
- Lifetime: a CA cannot issue certificates that outlive its own certificate, so the root must outlast every subordinate. Microsoft's example
CAPolicy.infsets a 20-year root renewal validity with a 4096-bit key, and recommends renewing a CA certificate once half its validity has elapsed. - Name it once: an AD CS CA cannot be renamed after installation without invalidating issued certificates. Do not use the server's FQDN or NetBIOS name as the CA common name.
- CRL only, no delta: an offline root publishes a full CRL on a long interval (for example six to twelve months) and has no delta CRLs, per Microsoft's guidance that deltas are unnecessary on offline roots. Diarise the CRL re-signing date: an expired root CRL breaks validation of every chain beneath it.
- Templates off: set
LoadDefaultTemplates=0on issuing CAs so no template is published until you choose it.
Designing issuing and policy CAs
| Issuing CA | Typical certificates | Key constraints | Online? |
|---|---|---|---|
| Directory authentication CA | Domain controllers, smart card and Windows Hello for Business, user and computer auto-enrolment | Client and server authentication EKUs; tight template ACLs | Yes, domain-joined, treated as Tier 0 |
| Network device CA | Wi-Fi (802.1X), VPN, NAC, device certificates enrolled by SCEP or EST | Client authentication only; device name constraints | Yes |
| TLS server CA | Internal web servers and APIs, often via ACME | Server authentication EKU; DNS name constraints to internal zones | Yes |
| Workload identity CA | Service mesh, SPIFFE, mTLS between services | Short lifetimes, URI SAN constraints | Yes, often automated or cloud-hosted |
| Code signing CA | Internal software, scripts, drivers | Code signing EKU only; issuance with manual approval | Online but heavily restricted, or offline |
- Enterprise versus standalone: in AD CS, enterprise CAs use Active Directory for templates and auto-enrolment; standalone CAs do not depend on AD and suit the offline root and offline policy CAs.
- Basic constraints: set
cA=TRUEwith apathLenConstrainton every CA certificate: 1 on a policy CA in a three-tier design, 0 on an issuing CA so it cannot create further CAs. RFC 5280 requires that keyCertSign is only asserted when the cA flag is set, and recommends marking key usage critical. - Name constraints: permitted and excluded subtrees for DNS names, email domains or URIs limit what a subordinate can issue. This is the most effective containment for a CA that might be compromised, as long as relying parties enforce it (test older devices).
- EKU on the CA certificate: many validators treat EKUs on intermediate CA certificates as a constraint on the chain. Put only the EKUs that branch needs.
- Policy OIDs: register an enterprise OID arc and assign one certificate policy OID per assurance level; put it in the
certificatePoliciesextension of the subordinate and end-entity certificates. RFC 5280 allows the specialanyPolicyOID (2.5.29.32.0) for CAs that do not restrict policy; avoid it below the root if you want policy to mean anything.
The Microsoft Storm-0558 Key Breach is not a CA compromise, but it illustrates the design point: a signing key that should have been retired stayed trusted, and relying services did not check which scope the key was meant for. Constraints and policy OIDs are how a PKI encodes scope so relying parties can check it.
Revocation and publication
- Decide AIA and CDP before first issuance: Microsoft notes that changing these URLs only affects newly issued certificates. Unreachable locations break VPN, smart card logon and email signature validation.
- HTTP first: publish CRLs and CA certificates on a highly available HTTP location reachable by every relying party, including non-Windows clients and external users. LDAP-only CDPs fail outside the domain. Windows tries CDP URLs in order, so put the most reachable first.
- OCSP for online CAs: run OCSP responders for issuing CAs with large CRLs or tight revocation needs, with their own short-lived signing certificates. See the Certificate Revocation Guide.
- Size CRL periods: issuing CA CRLs typically refresh daily to weekly with an overlap period; offline CA CRLs on a long interval with a hard calendar entry for renewal.
- Short-lived certificates: for workloads, lifetimes of hours or days can replace revocation checking. That moves the risk to the automation that renews them.
Certificate policy and CPS for identity use
RFC 3647 separates the certificate policy, which states what participants must do, from the certification practice statement, which describes how a specific CA meets those requirements. A CP can cover several CAs; a CPS is specific to one CA or operator and is more detailed. Both follow the same nine-part outline.
| RFC 3647 section | What an enterprise PKI should state |
|---|---|
| 1. Introduction | Hierarchy diagram, CA names, policy OIDs, permitted certificate uses |
| 2. Publication and Repository | AIA, CDP and OCSP URLs; who publishes; frequency |
| 3. Identification and Authentication | How a device, workload or person proves identity before issuance (auto-enrolment, attestation, manager approval) |
| 4. Certificate Life-Cycle Operational Requirements | Request, approval, renewal, re-key, revocation reasons and turnaround |
| 5. Facilities, Management, and Operational Controls | Physical security of the root, roles and dual control, logging, key ceremony records, business continuity |
| 6. Technical Security Controls | Key sizes, HSM validation level, key backup, activation data, CA lifetimes |
| 7. Certificate, CRL, and OCSP Profile | Extensions per certificate type, constraints, CRL validity |
| 8. Compliance Audit | Internal audit frequency, scope, who audits |
| 9. Other Business and Legal Matters | Liability to partners, data protection, change control of the CP |
For machine identities, section 3 is the one that matters most and is usually missing: state what evidence (TPM attestation, MDM enrolment, workload attestation, CI pipeline identity) is required before a certificate is issued. See the ISO/IEC 27099 PKI Practices and Policy Framework Guide for the ISO view.
Where identity fits
An internal CA is an identity provider in its own right. Every certificate it issues is a non-human credential, and in Active Directory a certificate with client authentication EKU can log on as the account named in it. SpecterOps' "Certified Pre-Owned" research (2021) catalogued template and CA misconfigurations (ESC1 to ESC8) that let low-privileged users obtain certificates for other accounts, and showed that a stolen CA private key lets an attacker forge certificates for any principal that cannot be revoked through normal means (ATT&CK T1649 Steal or Forge Authentication Certificates). Its recommendations: treat CA servers as Tier 0, audit published templates and protect CA keys in hardware. See the AD CS Security Guide and the Kerberos PKINIT and Certificate Mapping Guide.
Implementation approach
- Inventory first: list existing CAs, their trust distribution (GPO, MDM, OS stores), templates, issued certificate counts and every application that pins or trusts them.
- Write the CP and design: tiers, CA names, policy OIDs, constraints, lifetimes, AIA and CDP URLs, HSM model. Get it reviewed by the teams that will rely on it.
- Build in a lab: stand up the full hierarchy with test names, validate chains on every client type (Windows, macOS, Linux, network devices, mobile, Java trust stores) including name constraint enforcement.
- Root key ceremony: scripted, with at least two custodians, a witness, video or signed log, HSM backup to a second device and split activation credentials stored separately.
- Issue subordinates and publish: sign subordinate certificates, publish root and intermediates to trust stores and the HTTP repository, publish the first CRLs.
- Pilot: move one certificate type (for example Wi-Fi) to the new issuing CA for a pilot group; watch validation failures and enrolment errors.
- Migrate and run in parallel: keep the old hierarchy trusted until its last certificate expires; stop issuance from it first. Rollback is simple while the old CA still issues: repoint templates or enrolment profiles back.
- Decommission: revoke the old CA certificates, publish a final CRL, remove trust from stores, and destroy or archive keys per the CPS.
Decision guidance
- Self-run versus managed: running your own offline root gives full control but demands ceremonies, physical security and skills that are hard to keep in a small team. A managed or cloud private CA reduces operational risk; check whether keys are non-exportable, whether issuance is logged to your SIEM and whether you can enforce constraints.
- Small organisations: two tiers, one offline root, one or two issuing CAs, short lifetimes and automation. Do not build a three-tier hierarchy you cannot operate.
- Large organisations: three tiers where branches need independent revocation, dedicated issuing CAs per purpose, OCSP, and a PKI team with documented roles under the CPS.
- Public versus private: internal names and device identities belong on a private CA. Public CAs are for internet-facing TLS and are bound by CA/Browser Forum rules, including shrinking lifetimes; see the Certificate Expiry Outages and 47-Day TLS Certificates Guide.
Detection and measurement
- Log on issuing CAs: certificate requests, issuance, denials, template changes, CA configuration changes and private key use. In AD CS enable CA auditing so these appear in the Security log, and forward them to the SIEM.
- Watch for: certificates issued with a subject alternative name differing from the requester, new templates or template ACL changes, issuance outside business patterns, new root certificates appearing in endpoint trust stores (ATT&CK T1553.004 Install Root Certificate).
| Metric | Target |
|---|---|
| CA private keys held in a validated HSM | 100% for root and issuing CAs |
| Days until next offline CRL expiry | Never below the renewal lead time (for example 30 days) |
| CDP and AIA availability from outside the domain | 99.9% or better |
| Issuing CAs without path length and EKU constraints | 0 |
| Published templates reviewed in the last 12 months | 100% |
| Certificates issued by deprecated CAs still valid | Falling to 0 by the decommission date |
Framework mapping
| Framework | Reference | Relevance |
|---|---|---|
| NIST SP 800-53 Rev. 5 | SC-17, SC-12, SC-13, IA-5, CP-9 | PKI certificates, key management, cryptographic protection, authenticators, key backup |
| NIST SP 800-57 Part 1 Rev. 5 | Key management recommendations | Key types, protection and lifecycle for CA keys |
| NIST CSF 2.0 | PR.DS-01, PR.PS-01, ID.AM-08 | Protecting data at rest (keys), configuration management, lifecycle management |
| IETF | RFC 5280, RFC 3647 | Certificate and CRL profile; CP and CPS framework |
| MITRE ATT&CK | T1649, T1553.004, T1588.004 | Forging certificates, rogue roots, obtaining certificates |
Common failure patterns
- Online root that issues end-entity certificates: the root key sits on a domain-joined server with backup agents, and its compromise means rebuilding trust on every device.
- Software-protected CA keys: a domain admin or backup operator can export the issuing key and mint certificates indefinitely, the DPERSIST1 scenario from Certified Pre-Owned.
- Forgotten offline CRL: the root CRL expires, chains fail validation everywhere, and the root custodian has left the company.
- LDAP-only or internal-only CDP: remote laptops, network devices and Linux hosts cannot check revocation, so they either fail or skip checking.
- One issuing CA for everything: code signing, Wi-Fi and domain logon share a key, so a compromise of the least protected use forces revocation of all.
- No CP or CPS: nobody can say what proof a device needed to get a certificate, so audits and partner trust stall.
Practitioner checklist
- Document the hierarchy, CA names, policy OIDs and lifetimes in a CP before building anything.
- Keep the root offline, standalone and HSM-protected, with scripted dual-control ceremonies.
- Set path length, EKU and name constraints on every subordinate CA certificate.
- Separate issuing CAs by purpose, and treat AD-integrated CAs as Tier 0.
- Publish AIA and CDP over HTTP, reachable from every relying party, before first issuance.
- Diarise offline CRL renewals and CA certificate renewal at half-life.
- Move CA keys off FIPS 140-2-only modules at the next refresh, to FIPS 140-3 validated hardware.
- Enable CA auditing and alert on template, configuration and anomalous issuance events.
- Test the full chain, including name constraints and revocation, on every client type.
- Keep a tested runbook for revoking and replacing an issuing CA.
Standards and references
- IETF: RFC 5280 Internet X.509 PKI Certificate and CRL Profile
- IETF: RFC 3647 Certificate Policy and Certification Practices Framework
- NIST: SP 800-57 Part 1 Rev. 5 Recommendation for Key Management
- NIST: SP 800-53 Rev. 5
- NIST: Cryptographic Module Validation Program
- NIST: SP 1800-16 Securing Web Transactions, TLS Server Certificate Management
- Microsoft Learn: PKI design considerations using Active Directory Certificate Services
- SpecterOps: Certified Pre-Owned, Abusing Active Directory Certificate Services
Related NHI Mgmt Group resources: Machine Identity, PKI and Certificate Lifecycle Guide · AD CS Security Guide · ACME and Automated Certificate Management Guide · Certificate Revocation Guide · PKI Basics Guide