Skip to content

sec(iac): low-severity least-privilege gaps in hosted-infra grants (2026-09-02 audit) #323

Closed

Description

@cristim

Summary

Four low-severity grants in our own hosted infrastructure that are wider than the stated need, grouped because each is a one-resource Terraform edit by someone doing a least-privilege pass. None ships in a customer-deployed template. Two are latent (one because the role's Lambda cannot be applied today, one because the escalation needs a second WIF provider nobody has added); the other two are live but read-only over-grants.

Location

All at 3c0f8ac94048a2c36fce5ccddee54e6c4849a5cd:

  • terraform/environments/gcp/ci-cd-permissions/github_oidc.tf:46 (A13-014)
  • terraform/modules/secrets/aws/main.tf:390 (A13c-021)
  • terraform/modules/compute/gcp/cloud-run/main.tf:431 (A13c-022)
  • terraform/modules/compute/azure/container-apps/main.tf:269 (A13c-023)

Items

  • A13-014: GCP CI impersonation is granted to the whole repository on a generically named shared pool. The provider's attribute_condition pins repo and ref, but the workloadIdentityUser grant is principalSet://.../attribute.repository/<repo> with no ref component, and the pool is named github-actions. Principal identifiers are pool-scoped, so any second provider added to the pool that can mint attribute.repository for this repo satisfies the grant and impersonates a service account holding roles/resourcemanager.projectIamAdmin, roles/iam.roleAdmin, roles/cloudkms.admin and roles/storage.admin. iac/federation/gcp-target/terraform/main.tf:203-205 documents exactly this hazard for the customer module (and chore(terraform/environments): medium/low findings from the 2026-07-28 full review #132 item 8 tracks the customer-side instance); the CI pool has the same shape. PLAUSIBLE: with one provider in the pool this is defense in depth, not a reachable escalation.

    member = "principalSet://iam.googleapis.com/${google_iam_workload_identity_pool.github[0].name}/attribute.repository/${var.github_repo}"

    Fix: pin attribute.ref in the grant as well (a two-attribute member or condition) and give the pool a CUDly-scoped name.

  • A13c-021: the rotation Lambda's RDS grant wildcards the account and instance segments. Resource = "arn:aws:rds:${var.region}:*:db:*" authorises rds:ModifyDBInstance on any instance in the account, while the same resource's own count at :376 already holds the specific var.rds_cluster_id. Latent: the rotation Lambda this role serves cannot be applied today (see the placeholder-source item tracked in fix(iac/gcp): cleanup-function ships a placeholder.zip source that doesn't exist, breaking apply #124 / chore(terraform/modules): medium/low findings from the 2026-07-28 full review #131), so nothing assumes the role.

    Action = [
      "rds:ModifyDBInstance",
      "rds:DescribeDBInstances"
    ]
    Resource = "arn:aws:rds:${var.region}:*:db:*"

    Fix: build the ARN from data.aws_caller_identity.current.account_id and var.rds_cluster_id.

  • A13c-022: the Cloud Run service account holds project-wide roles/compute.viewer. The comment names only compute.commitments.list and compute.machineTypes.list as the need; roles/compute.viewer grants read on every Compute Engine resource in the project, including instance metadata and startup scripts. The module already defines a two-permission custom role for the write side immediately below (:454-463), so the narrow pattern is available.

    resource "google_project_iam_member" "compute_viewer" {
      project = var.project_id
      role    = "roles/compute.viewer"
      member  = "serviceAccount:${google_service_account.cloud_run.email}"
    }

    Fix: extend cudlyCommitmentWriter (or add a sibling custom role) with compute.commitments.list/get and compute.machineTypes.list/get, and drop the predefined role.

  • A13c-023: the Container App identity holds Reader over the entire host subscription. azurerm_role_assignment.subscription_reader has no count and is scoped to local.subscription_resource_id, the same subscription that holds CUDly's own Key Vault, PostgreSQL server, ACR and Container App environment, so a compromised runtime can enumerate the whole control plane of its own deployment. The stated need (host account ingested as a Self account) is an opt-in mode; the grant is unconditional.

    resource "azurerm_role_assignment" "subscription_reader" {
      scope                = local.subscription_resource_id
      role_definition_name = "Reader"
      principal_id         = azurerm_user_assigned_identity.container_app.principal_id
    }

    Fix: gate it behind an enable_self_account_ingestion input defaulting false.

Suggested fix

Four independent single-resource edits; each item carries its own.


Found by the 2026-09-02 codebase audit, finding A13-014, A13c-021, A13c-022, A13c-023, reported by one reviewer and independently confirmed by a second. Full report: docs/audits/codebase-audit-2026-09-02.md.

Activity

  1. IAmJSD commented on Oct 23, 2026

    @IAmJSD
    Collaborator

    Claiming: working on this

  2. self-assigned this
    on Oct 23, 2026
  3. cristim commented on Nov 12, 2026

    @cristim
    MemberAuthor

    Status on main (faf1bcb, claude-platform): A13c-023 is done: the host subscription Reader grant is opt-in behind enable_self_account_ingestion (default false). Existing Azure deployments lose that grant on the next apply unless the variable is set; Azure deploys are paused. A13c-021 and A13c-022 already hold on main (the RDS ARN is built from the account id; the GCP compute grant is a custom role with a guard test). A13-014 is partly done: the grant member already pins repo id and ref; the pool rename is not done, because pool IDs are ForceNew and a rename would destroy the pool and break CI auth. The issue stays open until the owner decides on the pool rename.

  4. cristim commented on Nov 14, 2026

    @cristim
    MemberAuthor

    Claimed by claude-platform-2 (coordinator assignment 2026-11-14, not touching store_postgres.go, plans.ts, recommendations.ts): premise-check against main, then implement and close (Closes #323), or comment the evidence/blocker.

  5. cristim commented on Nov 14, 2026

    @cristim
    MemberAuthor

    Closing: three of the four grants are already fixed on main: A13c-021 rotation Lambda RDS grant scoped to the cluster ARN (898a166, #1290), A13c-022 Cloud Run compute.viewer replaced by a custom role with only compute.commitments.get/list and compute.machineTypes.list (guard test terraform/modules/compute/gcp/runtime_compute_iam_test.go), A13c-023 Azure Container App Reader is opt-in via enable_self_account_ingestion (faf1bcb). A13-014 (GCP CI impersonation): the workloadIdentityUser member now pins repo id and ref (attribute.repository_id_ref, 63c464a/#1066) and the provider attribute_condition pins owner id, repo id and ref, with create_before_destroy so the upgrade has no access gap; that closes the escalation path the finding describes. Decided (owner rule): the remaining 'rename the github-actions pool' is defense in depth only and workload_identity_pool_id is ForceNew (replaces the pool and provider, breaks CI auth for every GCP workflow until a bootstrap re-apply and a GCP_WORKLOAD_IDENTITY_PROVIDER variable update), so it is not done; logged as an optional owner action.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions