Cloud pentest report: scope and coverage
Brief
The findings say what is wrong; the report around them says where we looked, as whom, for how long, what we never reached and how the client proves each fix. Every number in those sections comes from a command or a log the client can rerun, so a clean result cannot be read as a promise about parts of the estate nobody reviewed.
On this page
- Check first
- Which identity and accounts does the scope section name?
- How do I write the coverage table?
- Which services did the review actually touch?
- What goes in the statement of changes?
- How do I write the retest plan?
- What goes in the executive summary?
- The defender sees
- Done when
- Not covered
- Tools
- Related research
- Where this page fits
- Reviewed
- Situation
- Confirmed findings already written up as on the cloud finding page, the operator log and evidence package from the whole test, and the review identity still valid for a few more reads. Every command here reads CloudTrail, the account's region list or the policy simulator, and nothing changes the account. The report covers only the accounts, subscriptions and projects the scope letter names, and says so.
- Goal
- The sections around the findings: scope, coverage table, services reviewed, statement of changes, retest plan and the executive summary written last, each number traceable to a command.
- Outputs
- Representative examples on placeholder hosts, trimmed to the lines that matter. Not captured from a public fixture yet.
Check first
- Every finding is confirmed and written up, each with its own evidence pointer and retest command.
- The scope letter and the scope check record from the scoping page: the identifiers the report repeats come from there.
- The window in UTC, start and end, from the first and last operator log lines.
- The client's audience: an executive summary for whoever signs off, the rest for the platform team that fixes.
- The evidence package is hashed, as on the evidence package page, so the report can cite the manifest hash.
Steps
Which identity and accounts does the scope section name?
aws sts get-caller-identity --query '[Account,Arn]' --output text
Expected output
111122223333 arn:aws:sts::111122223333:assumed-role/atarus-review/atarus-review # <- the identity every claim in the report was read as
The scope section names each account, tenant, subscription and project by identifier, the identity used in each with its role, and the window. It is read fresh on the day of writing, because a role renamed or a session name changed since the test would break every evidence pointer. On Azure the same line comes from az account show, on Google Cloud from gcloud config list, both shown on the scoping page. Anything the letter named that we could not reach, such as a subscription the review principal never saw, is listed here with the reason, not left out.
Negative result. The identity differs from the one in the operator log: say which identity read what, account by account, before anything else in the report.
Fix. Keep review roles in place until the report is accepted, then remove them; the report's last appendix records that removal.
How do I write the coverage table?
printf 'enabled\t%s\n' "$(aws account list-regions --region-opt-status-contains ENABLED ENABLED_BY_DEFAULT \
--query 'Regions[].RegionName' --output text | wc -w | tr -d ' ')"
printf 'reviewed\t%s\n' "us-east-1 eu-west-1"
Expected output
enabled 18
reviewed us-east-1 eu-west-1 # <- two of eighteen: the table says so in so many words
Coverage, 2 to 6 Nov 2026, read-only role atarus-review
Account Regions enabled Reviewed Services reviewed Not reviewed
111122223333 18 us-east-1, eu-west-1 IAM, STS, S3, KMS, Lambda, 16 other regions; EKS workloads;
EC2, SSM, CloudTrail data in any bucket
Subscription n/a eastus RBAC, managed identities, Key Vault contents; AKS
corp-prod storage, Activity Log
The table is the line that stops a clean result reading as "nothing wrong anywhere". It goes account by account and states what was not reviewed as plainly as what was, including what the rules excluded on purpose. wc -w counts the region names; tr strips the padding BSD wc adds on macOS. Single findings carry their own coverage line, as on the finding page; this table is the sum for the whole test.
Negative result. Every enabled region and every service the letter named was reviewed; write that, with the date, and it is still a statement about those accounts only.
Fix. Deny unused regions with an SCP on aws:RequestedRegion and leave opt-in regions off, so the table can say where nothing can run rather than where nobody looked.
Which services did the review actually touch?
aws cloudtrail lookup-events --region us-east-1 --start-time 2026-11-02T14:00Z --end-time 2026-11-06T22:00Z \
--lookup-attributes AttributeKey=Username,AttributeValue=atarus-review \
| jq -r '.Events[].CloudTrailEvent | fromjson | .eventSource' | sort | uniq -c | sort -rn
Expected output
412 iam.amazonaws.com
118 s3.amazonaws.com
64 kms.amazonaws.com
51 lambda.amazonaws.com
37 ec2.amazonaws.com
22 sts.amazonaws.com
9 cloudtrail.amazonaws.com
4 ssm.amazonaws.com # <- a service named in the scope with only four calls: check it was covered, not glanced at
The services list in the report comes from the review's own events, not from memory. lookup-events reads one region's event history at a time, management events only, for 90 days, so the command runs once per region reviewed; IAM and global STS calls land in us-east-1. For an assumed role, the Username attribute matches the role session name, which is why the review role's session name is fixed in the letter. A service the coverage table claims with no events behind it is either a mistake in the table or a service reviewed through another identity, and the report says which.
Negative result. Every service in the coverage table has events under the review identity in each region it claims.
Fix. Keep management events from every region in an organization trail, so this count is possible after the 90 days of event history run out.
What goes in the statement of changes?
aws cloudtrail lookup-events --region us-east-1 --start-time 2026-11-02T14:00Z \
--lookup-attributes AttributeKey=ReadOnly,AttributeValue=false \
| jq -r '.Events[] | select((.Username // "") | startswith("atarus")) | .CloudTrailEvent | fromjson
| [.eventTime, .eventName, (.errorCode // "-")] | @tsv'
Expected output
2026-11-04T15:22:40Z DeleteTags -
2026-11-04T15:21:02Z CreateTags - # <- the one change the letter named, and its removal
2026-11-03T10:08:13Z CreateTags Client.DryRunOperation
The statement of changes lists every write the test made, with its clean-up, and nothing else. It is checked against write events rather than written from the operator log, because the log records what we meant to do and CloudTrail records what happened. select keeps events whose user name starts with our session prefix and skips events that carry no user name, which an expression that assumes one would fail on. A dry run appears as a write event with its error code, and the report lists it as attempted, not made.
Negative result. No write events at all: the statement reads "no changes were made to any account in scope", and it now has evidence behind it.
Fix. Alert on any write by an identity named for review, since a review role that can write is a finding about the role.
How do I write the retest plan?
sim() {
printf '%s\t%s\t' "$1" "$2"
aws iam simulate-principal-policy --policy-source-arn arn:aws:iam::111122223333:user/ci-deploy \
--action-names "$1" --resource-arns "$2" \
--query 'EvaluationResults[0].EvalDecision' --output text
}
sim iam:CreatePolicyVersion arn:aws:iam::111122223333:policy/ci-deploy-policy
sim iam:PassRole arn:aws:iam::111122223333:role/app-deploy-role
sim lambda:CreateFunction arn:aws:lambda:us-east-1:111122223333:function:atarus-passrole-proof
sim sts:AssumeRole arn:aws:iam::111122223333:role/ops-break-glass
sim ssm:SendCommand arn:aws:ec2:us-east-1:111122223333:instance/i-0abc1234def567890
sim ssm:SendCommand arn:aws:ssm:us-east-1::document/AWS-RunShellScript
Expected output
iam:CreatePolicyVersion arn:aws:iam::111122223333:policy/ci-deploy-policy allowed # <- finding 1, expected implicitDeny after the fix
iam:PassRole arn:aws:iam::111122223333:role/app-deploy-role allowed
lambda:CreateFunction arn:aws:lambda:us-east-1:111122223333:function:atarus-passrole-proof allowed
sts:AssumeRole arn:aws:iam::111122223333:role/ops-break-glass allowed
ssm:SendCommand arn:aws:ec2:us-east-1:111122223333:instance/i-0abc1234def567890 allowed
ssm:SendCommand arn:aws:ssm:us-east-1::document/AWS-RunShellScript allowed # <- a send needs both the instance and the document
Retest plan
Finding Check Today Expected after fix
1 simulate ci-deploy iam:CreatePolicyVersion on ci-deploy-policy allowed implicitDeny
2 simulate ci-deploy iam:PassRole on app-deploy-role allowed implicitDeny
3 get-snapshot-block-public-access-state us-east-1 unblocked block-all-sharing
4 az role assignment list, Owner at subscription 3 rows 1 row (break-glass)
One line per finding, each a command the client can run themselves with today's result and the result that closes it. Where several findings rest on the same principal, one simulation step covers them all and becomes the first retest, with one call per action that names the resource its finding names. One resource goes in each call: with two named, the simulator returns one result per action, whose EvalDecision is the most restrictive of them and whose EvalResourceName is a template rather than either ARN, so sim prints the action and resource it was given beside the decision. A simulation that names no resource tests every action against *, so a grant written for one role, policy or instance prints implicitDeny before the fix as well as after it, and the retest would close a finding nobody fixed. The plan also says what a retest does not prove: a simulator implicitDeny shows the identity policy changed, not that a resource policy elsewhere stopped granting the same action, so findings that rest on a resource policy are retested by reading that policy.
Negative result. A finding with no command that shows it fixed: rewrite the finding until it has one.
Fix. Put the retest commands into the client's pipeline as policy tests, so a fixed line cannot come back unnoticed.
What goes in the executive summary?
Summary. Between 2 and 6 Nov 2026 we reviewed AWS account 111122223333 (two of eighteen
enabled regions) and Azure subscription corp-prod with read-only identities the client
issued. One finding is Critical: the CI deployment key can make itself an administrator
in four ways, each shown by simulation and one by a proof removed the same hour. Two are
High and four Medium. We made one change, a tag, and removed it; CloudTrail recorded
every step and nothing alerted. The Critical closes with one policy change, and every
finding has a one-command retest in section 6.
The summary is written last and read first. It gives the scope in one sentence, the worst finding in plain words with how it was shown, the counts, the statement of changes, the defender's view in one clause and the first fix, in that order. It names no person and makes no claim about anything outside the coverage table.
Negative result. A summary that could be pasted into any report: it lacks an account, a count or a date, and needs rewriting.
Fix. The client reads the summary against the coverage table before sign-off; anything it claims that the table does not support comes out.
The defender sees
The reads that fill these sections are in CloudTrail under the review session: GetCallerIdentity, ListRegions, LookupEvents and SimulatePrincipalPolicy. Their place in the report is the defender's view section, written as on the CloudTrail and GuardDuty evidence review and the Entra sign-in and audit log review, with Azure control-plane writes from the Activity Log review. The report says what was recorded, what was never logged, what alerted and who acted, about our activity only, and never claims the accounts are free of anyone else's.
Done when
Scope names every identity and identifier used, coverage gives enabled against reviewed regions per account, the services list comes from the review's own events, the statement of changes is backed by a write-event check, every finding has a retest line with its expected result, and the summary is written from those sections.
Not covered
Writing the individual finding (see how to write up a cloud finding), severity scoring in general (see writing a finding), compliance mappings and pricing of the fixes; the outputs shown are representative examples on placeholder accounts, not captured runs.