AWS Organizations and SCP guardrail review
Brief
Service control policies are the ceiling on what any identity in a member account can do, so a weak or missing guardrail turns every escalation path in that account into an organization problem. The review reads the policies, where they attach, how root users are handled and which member accounts hold organization-wide power, all from a read-only role.
On this page
- Situation
- A read-only review role in the management account, or in the member account the client has made delegated administrator for Organizations, with
AWSOrganizationsReadOnlyAccessand IAM read. Every command on this page is a read; we never attach, detach or edit a policy, never move an account and never start a root session, and we review only the organization the scope letter names. - Goal
- A guardrail map: each SCP with what it denies and where it attaches, the accounts no guardrail reaches, the root user controls for member accounts, and every delegated administrator with its services.
- Outputs
- Captured on 8 Nov 2026 (UTC) in the Atarus lab AWS Organization from the Atarus Bench edition
2026.11-cloud-0.3, AWS CLI 2.39.3 and jq 1.7.1, with account ids, policy ids, OU ids and other identifiers replaced by stable placeholders. Every step on this page is captured: the Organizations and IAM reads through a read-only role in the lab's management account, which runs no workload, and the account summary through the review role in member account111122223333. The lab Organization has trusted access for IAM off, so the root access check shows the off state as an error. The logging account's delegated services come from a second run that day.
Check first
- The management account id, and whether the review role reads Organizations from there or from a delegated administrator.
- The client's organizational units and which accounts sit in each, from their own account list.
- The guardrails the client believes it has, written down before we read a single policy.
- Whether centralized root access is turned on for member accounts.
- Where the organization trail delivers, so each read below can be found again.
Steps
AWS
Are service control policies on, and what do they deny?
aws organizations describe-organization --query 'Organization.[MasterAccountId,FeatureSet]' --output text
aws organizations list-roots --query 'Roots[].PolicyTypes[].[Type,Status]' --output text
aws organizations list-policies --filter SERVICE_CONTROL_POLICY \
--query 'Policies[].[Id,Name,AwsManaged]' --output text
aws organizations describe-policy --policy-id p-guard01 --query Policy.Content --output text \
| jq -r '.Statement[] | select(.Effect == "Deny") | [.Sid, ([.Action] | flatten | join(","))] | @tsv'
Expected output
123456789012 ALL
SERVICE_CONTROL_POLICY ENABLED
p-aaaa0001 DenyLeaveAndCloseAccount False
p-guard01 lab-guardrail False
p-FullAWSAccess FullAWSAccess True
p-aaaa0002 lab-allow False
p-aaaa0003 lab-spend-stop False
p-aaaa0004 capture-allow False
DenyOtherRegions # <- an empty action column: written with NotAction, so read it in full
DenyLargeInstances ec2:RunInstances
DenyTrailChanges cloudtrail:StopLogging,cloudtrail:DeleteTrail,cloudtrail:PutEventSelectors,cloudtrail:UpdateTrail
DenyEmailSending ses:*
DenyLeavingOrg organizations:LeaveOrganization
SCPs grant nothing. They set the most any identity in a member account can do, root user included, so the useful question is what they deny. A feature set of CONSOLIDATED_BILLING means SCPs cannot be used at all. FullAWSAccess is the default that allows everything; the guardrails are the deny statements beside it. A region deny is usually written with NotAction, which prints an empty action column here, so we read any such statement in full, as DenyOtherRegions is here. The lab's guardrail denies other regions, large instances, changes to the trail, sending email and leaving the organization, and nothing in it protects GuardDuty, Config or the platform team's roles. The step reads one policy; DenyLeaveAndCloseAccount and lab-spend-stop are read the same way, and lab-allow and capture-allow are allow lists, which the next step places. We compare the result with the list the client wrote down.
Negative result. Feature set ALL, SCPs enabled, and deny statements covering leaving the organization, turning off CloudTrail, GuardDuty and Config, changing the platform team's roles, and the regions the client does not use.
Fix. Write the guardrails as deny statements in a few SCPs, with an exception condition on aws:PrincipalArn naming only the platform role, and test each one in a sandbox OU before attaching it widely.
Which accounts sit outside the guardrails?
for t in r-ab12 $(aws organizations list-organizational-units-for-parent --parent-id r-ab12 \
--query 'OrganizationalUnits[].Id' --output text); do
printf '%s\t%s\n' "$t" "$(aws organizations list-policies-for-target --target-id "$t" \
--filter SERVICE_CONTROL_POLICY --query 'Policies[].Name' --output text)"
done
aws organizations list-accounts-for-parent --parent-id r-ab12 --query 'Accounts[].[Id,Name]' --output text
Expected output
r-ab12 DenyLeaveAndCloseAccount lab-guardrail FullAWSAccess # <- the guardrail attaches at the root, so it reaches every member account
ou-ab12-aaaa0001 FullAWSAccess capture-allow
ou-ab12-aaaa0002 lab-allow # <- an allow list in place of FullAWSAccess: accounts here can do only what it allows
999988887777 atarus-lab-logging
123456789012 atarus-management
444455556666 capture-b # <- a member account directly under the root: the root's policies reach it, no OU's do
An SCP applies to the target it is attached to and everything below it, so a guardrail on an OU reaches only the accounts under that OU. Accounts directly under the root get only what the root carries; where the root carries nothing but FullAWSAccess, they and any OU nobody attached a policy to run with no guardrail at all. An action must also be allowed at every level from the root down, so an OU whose only policy is an allow list, as the second OU here is, limits its accounts to that list. Nested OUs need the same loop one level down. r-ab12 is the root's id, the Id field of list-roots; the query in the step above prints the root's policy types instead. In the lab the guardrail sits at the root and reaches every member account, but two members, the logging account and capture-b, sit directly under the root, which the clean line below does not allow: OU policies, such as the allow lists here, never reach them, and a guardrail added at an OU later would miss them. SCPs never apply to the management account at all, and they do not restrict service-linked roles, so the management account must hold no workloads and very few people.
Negative result. The baseline guardrails attach at the root or at every top-level OU, no member account sits directly under the root, and the management account runs nothing but organization tooling.
Fix. Attach the baseline guardrails at the root, move stray accounts into an OU, and keep workloads and day-to-day users out of the management account.
Are root users controlled from the organization?
aws iam list-organizations-features
aws iam get-account-summary --query 'SummaryMap.[AccountMFAEnabled,AccountAccessKeysPresent]' --output text
Expected output
aws: [ERROR]: An error occurred (ServiceAccessNotEnabledException) when calling the ListOrganizationsFeatures operation: Trusted Access for IAM not enabled by organization of input account 123456789012. # <- centralized root access is off: each member account keeps its own root credentials
0 0 # <- in 111122223333: no root MFA, and no root access key
The first call runs in the management account; the second runs in each member account, through the review role there. list-organizations-features answers only once the organization has enabled trusted access for IAM, which centralized root access needs, so ServiceAccessNotEnabledException is the answer: the features are off. With centralized root access, the management account can remove root passwords, access keys and MFA devices from member accounts and run short root tasks itself when one is needed. Without it, every member account root is one password reset away from whoever reads the root email address, which is why 0 under MFA is the finding here even with no access key. We read the feature state and the account summary only; we never start a root session.
Negative result. RootCredentialsManagement and RootSessions both enabled, member accounts with no root credentials, or where it is off, MFA on every root user and no root access keys.
Fix. Enable both features, delete member account root credentials, deny root user actions in an SCP, and make every root email address a list the security team reads.
Who holds delegated administrator rights?
aws organizations list-delegated-administrators --query 'DelegatedAdministrators[].[Id,Name]' --output text
aws organizations list-delegated-services-for-account --account-id 999988887777 \
--query 'DelegatedServices[].ServicePrincipal' --output text
aws organizations list-delegated-services-for-account --account-id 111122223333 \
--query 'DelegatedServices[].ServicePrincipal' --output text
Expected output
999988887777 atarus-lab-logging # <- the one delegated administrator, a member account directly under the root
access-analyzer.amazonaws.com # <- it administers IAM Access Analyzer for the organization
aws: [ERROR]: An error occurred (AccountNotRegisteredException) when calling the ListDelegatedServicesForAccount operation: The provided account is not a registered delegated administrator for your organization. # <- a workload account that holds no delegation
The second call lists what the delegated administrator holds: IAM Access Analyzer, one service in a dedicated logging account, which is what a client would expect, though the account sits directly under the root rather than in a guarded OU. The third asks the same of the workload account 111122223333 and is refused with AccountNotRegisteredException, the clean result for a workload account. Delegation moves organization-wide power into a member account. A StackSets delegated administrator can deploy to every member account, and an IAM Identity Center delegated administrator can assign permission sets in them. Anyone who is administrator in that account holds that reach. SCPs do apply to delegated administrators, so one in an account no guardrail reaches holds organization-wide power with nothing over it.
Negative result. Delegated administrators only in dedicated security or identity accounts under a guarded OU, each holding only the services the client expects.
Fix. Remove delegations nobody uses, keep the rest in a dedicated account under the guardrails, and control administrator access there as tightly as the management account.
The defender sees
Organizations is a global service, so every Organizations call above is recorded by CloudTrail in us-east-1 of the management account under the review role: DescribeOrganization, ListRoots, ListPolicies, DescribePolicy, ListOrganizationalUnitsForParent, ListPoliciesForTarget, ListAccountsForParent, ListDelegatedAdministrators and ListDelegatedServicesForAccount, with IAM's ListOrganizationsFeatures. GetAccountSummary is recorded in the member account it ran in, under that account's review role. A role reading every SCP is normal during a review and worth a question at any other time. The events that matter more are changes: DetachPolicy, UpdatePolicy, MoveAccount, RegisterDelegatedAdministrator and DisableAWSServiceAccess should each raise an alert. CloudTrail takes minutes to make an event searchable, so a search straight after the review can come back empty for that reason alone. If the client still cannot find our reads in the organization trail a quarter of an hour later, that is the finding, and the cloud first hour covers the logging checks inside each account.
Done when
Every SCP is read and summarised, each root and top-level OU has its attached policies recorded, every account outside a guardrail is named, root credential status is known, and each delegated administrator is listed with the services it holds.
Not covered
Escalation inside one account (see AWS IAM privilege escalation checks), resource control policies in depth, tag and backup policies, and Control Tower drift.