ADR-0020: Azure subscription access through three groups
Date: 2026-09-18 Status: proposed Decider: platform owner
Context
Read live on 2026-09-15 and again on 2026-09-18, the tenant's only subscription carried 22 role assignments, every one made directly to a person or a service principal. Ten people held Owner. No group was used for access. Three different people had made the grants over four years, so the only record of why anyone holds a role is the assignment's own timestamp. One person held no role at all and could not see the subscription; two people each held two accounts, both privileged. The tenant holds no Entra licence, so Privileged Identity Management and Conditional Access are unavailable, and nine of the ten Owners sign in with an identity this tenant cannot put MFA on (personal Microsoft accounts and e-mail one-time passcodes).
ADR-0019 puts production in its own subscription. That subscription is about to exist. The access model has to be decided before it does, or production inherits the same pattern on day one. The staging and production pilot plan routes the subscription as item STP-P1-10 and the trust-root grants on it as STP-P1-11.
Decision
People hold Azure roles only through three Entra security groups. Each group binds one role at one scope. Nobody holds a direct assignment except one anchor.
| Group | Role | Control-plane subscription | Production subscription |
|---|---|---|---|
azure-owners |
Owner | bound | bound |
azure-contributors |
Contributor | bound | not bound |
azure-readers |
Reader | bound | bound |
Six rules follow.
- One door. A person gets access by one row in the roster that lives in the trust-root stack, in a pull request the owner merges, applied by a person through the trust root's break-glass window. There is no other route: a grant made in the portal is drift.
- Roles by job, not by person. Three roles exist for people. A need that none of them fits is a new decision, not a fourth direct grant.
- Groups hold roles; people hold memberships. Adding or removing a person changes a group, never a role binding.
- Least standing privilege.
azure-ownersholds at most two named people. Production has no human Contributor at all: CI applies production through the bound two-phase dispatch, and people read it. - Every grant is in git. The roster row records the object id, the group, the date and the reason. The pull request records who approved it.
- One anchor, outside the model. The platform owner's account keeps its direct Owner and its Global Administrator role, and no automation manages that assignment. It is the recovery path if a change to the groups goes wrong.
The bindings live in the trust-root stack and are applied by a person because the CI apply identity may not grant roles to users or groups: its RBAC Administrator role is constrained by an ABAC condition to service principals only (trust-root hardening, 2026-09-15). That constraint is kept. Roster changes therefore cost a break-glass window each, which is acceptable at the current rate of joiners and keeps human access to the control-plane subscription in human hands.
The control-plane subscription (dev, staging, the shared registry, the DNS zone, the Terraform state account, the Argo CD hub) is part of production's trust boundary: whoever controls it controls what production runs. Reducing its Owners is therefore part of going to production, not housekeeping.
Options considered
Option A — three groups, bindings in the trust root, human-applied (chosen)
- Benefit. One place to look, one place to change, a fixed role set, an audit trail in pull requests, and the same model on the production subscription from its first day. No spend.
- Cost or risk. Each roster change is a break-glass window until membership moves to a CI-applied root. The window script proves group access is effective before it removes a direct grant.
- Rollback or migration effect. The anchor is never touched; the bindings are additive; removing a group binding restores nothing that was lost, so removal is an owner deletion decision.
Option B — keep direct assignments, add a periodic audit
- Benefit. No ceremony.
- Cost or risk. The audit finds the mess after the fact and cannot say why a grant exists. It is what produced ten Owners.
- Rollback or migration effect. None; this is the state being left.
Option C — Entra ID P2 with Privileged Identity Management
- Benefit. Time-bound, approved elevation; access reviews.
- Cost or risk. A per-user licence for every privileged person, and a Conditional Access design before it is worth anything. Not justified while the privileged set is a handful of people and no identity here is MFA-enforceable.
- Rollback or migration effect. Can be layered on Option A later: groups become PIM-eligible without changing the role bindings.
Consequences
- Implementation work. Two break-glass windows: one creates the groups, binds the roles, populates the roster and removes every other direct human assignment at subscription scope together with a legacy service principal that held Contributor with an expired secret; the other creates the production subscription and binds the groups and the CI apply identity on it. A runbook for adding a person mirrors the existing promoter runbook.
- Tests and operating work. Each window ends with a readback: the subscription scope must hold exactly the anchor, the three group bindings and the CI identities, and every roster member must show the role through the group. The daily inventory snapshot records the resulting state.
- Remaining risks, each a follow-up and none a blocker for production. A Contributor on the control-plane subscription can still reach production through the Argo CD hub, the shared registry, or by re-enabling shared-key access on the state account: a deny policy on shared-key access and a pruned Contributor roster close that. Most people still sign in with identities this tenant cannot enforce MFA on: re-inviting them from the organisation's own Entra tenant with cross-tenant MFA trust closes that. The anchor is a single Global Administrator: a sealed emergency-access account closes that. A management group above both subscriptions carries the bindings and Azure Policy once, so a third subscription inherits them.
Supersession and rollback
A later ADR may move roster membership into a CI-applied root (the apply identity as group owner with the Group.Create application role), place the bindings on a management group, or adopt Privileged Identity Management. Each keeps the three groups. Deleting a binding or a group is an owner deletion decision, never an implementation rollback.
Program links
- Program decision: STP-D10 (subscription ownership), through ADR-0019
- Dependent work packages: STP-P1-10, STP-P1-11; roadmap line "RBAC cleanup from the July access audit"