Prove cloud impact without causing damage
Brief
Impact in a cloud account is proven by showing what an identity could do, not by doing it: one object read and hashed, a simulated delete, one tag written and removed, and a dry run priced from the public pricing page. The client is left with two log lines to read instead of a deleted database or a bill.
On this page
- Check first
- How do I prove read access to data without copying it?
- How do I show a destructive permission without using it?
- How do I prove write access with one tag?
- How do I show a cost impact without running anything?
- Did we change only what we said we would?
- The defender sees
- Done when
- Not covered
- Tools
- Related research
- Where this page fits
- Reviewed
- Situation
- A finding from an earlier stage, such as an access key that reaches a production account, and a statement of what the client wants proven. The rules of engagement name the one object we may read and the one resource we may tag; everything else on this page is a read, a simulation or a dry run, and only the accounts and regions the scope letter names are touched.
- Goal
- An evidence set that proves what the identity could reach and change, with one tag, written and removed, as the only change made.
- Outputs
- Representative examples on placeholder hosts, trimmed to the lines that matter. Not captured from a public fixture yet.
Check first
- The rules of engagement name the one object we may read and the one resource we may tag, and say nothing else may be written.
- The client named what they want proven: a bucket of customer exports, a production database, the bill a stolen key could run up. We prove that and stop.
- The identity we act as, and a separate review role that can run the IAM policy simulator and read CloudTrail.
- No automation keys on the tag name we plan to use, so writing it changes nothing a workload reads.
- A client contact we call before any step that could touch production in a way nobody expected.
Steps
How do I prove read access to data without copying it?
aws s3 ls s3://exports-prod-example/ --recursive --summarize | tail -n 2
aws s3api head-object --bucket exports-prod-example --key atarus-impact-check.txt \
--query '[ContentLength,ServerSideEncryption,SSEKMSKeyId]' --output text
aws s3 cp s3://exports-prod-example/atarus-impact-check.txt - | shasum -a 256
Expected output
Total Objects: 18342
Total Size: 41873520113
48 aws:kms arn:aws:kms:us-east-1:111122223333:key/1a2b3c4d-5e6f-4a5b-8c7d-9e0f1a2b3c4d
9f4e2a71c08d5b36e1a7f2c94d0b6e83a5c17f29d4e08b61c3a9f5d27e0b4c18 - # <- read and decrypted, never written to disk
The count and size prove the identity can list everything. The one object proves it can read and, because the object is encrypted with a customer key, decrypt. We ask the client to place that object for the test, or to name one that holds no personal data, and we keep its hash instead of its contents. The report says "the key could list 18,342 objects totalling 39 GiB and read them", and stops there.
Negative result. AccessDenied on the listing or on GetObject, or a KMS.AccessDeniedException on the copy, which proves the key policy held even where the bucket policy did not.
Fix. Scope bucket access to the roles that need it, encrypt sensitive buckets with a customer managed key whose policy names those same roles, and log S3 data events on buckets that hold customer data.
How do I show a destructive permission without using it?
sim() {
aws iam simulate-principal-policy --policy-source-arn arn:aws:iam::111122223333:role/ci-deploy \
--action-names "$1" --resource-arns "$2" \
--query 'EvaluationResults[].[EvalActionName,EvalDecision]' --output text
}
sim s3:DeleteObject arn:aws:s3:::exports-prod-example/export-2026-06.csv
sim s3:PutBucketPolicy arn:aws:s3:::exports-prod-example
sim kms:ScheduleKeyDeletion arn:aws:kms:us-east-1:111122223333:key/1a2b3c4d-5e6f-4a5b-8c7d-9e0f1a2b3c4d
sim rds:DeleteDBInstance arn:aws:rds:us-east-1:111122223333:db:orders-db
Expected output
s3:DeleteObject allowed
s3:PutBucketPolicy allowed
kms:ScheduleKeyDeletion implicitDeny # <- denied on the key itself
rds:DeleteDBInstance allowed # <- could delete the production database; nothing was called
The simulator runs from our review role and evaluates the identity's own policies, so nothing is attempted against a real resource. Each call names the resource the client named, one action to a call, since the simulator pairs every action with every resource it is given. Without --resource-arns it tests each action against *, and a grant written for a named bucket, key or database never matches *, so the delete the client fears most could come back implicitDeny while the identity holds it (the rule is on the policy evaluation page). The simulator does not see a resource's policy unless we pass one, so the report says "allowed by the role's policies" and names any resource policy that would still stop it.
Negative result. implicitDeny or explicitDeny on every destructive action the client named, each simulated against the resource it would act on.
Fix. Remove delete and policy-change actions from roles that only deploy, turn on deletion protection for databases, and deny kms:ScheduleKeyDeletion outside the platform role with an SCP.
How do I prove write access with one tag?
aws ec2 create-tags --region us-east-1 --resources i-0abc12de34f567890 --tags Key=atarus-test,Value=2026-10-28
aws ec2 describe-tags --region us-east-1 --filters Name=resource-id,Values=i-0abc12de34f567890 Name=key,Values=atarus-test --output text
aws ec2 delete-tags --region us-east-1 --resources i-0abc12de34f567890 --tags Key=atarus-test
Expected output
TAGS atarus-test i-0abc12de34f567890 instance 2026-10-28 # <- written, read back, then removed
create-tags adds one key and leaves the instance's other tags alone, and delete-tags removes only that key. On Azure, az tag update --operation Merge and --operation Delete do the same. A tag changes no data and no running process, yet the permission that wrote it is the one that matters, so it proves the write without the risk.
Negative result. UnauthorizedOperation on CreateTags. Report the read access and stop; do not look for another way to write.
Fix. Grant tag writes only to the deployment role, and where tags drive access decisions, deny tag changes to everyone else with an SCP.
How do I show a cost impact without running anything?
aws ec2 run-instances --region us-east-1 --dry-run --image-id ami-0123456789abcdef0 --instance-type m5.xlarge --count 1
aws service-quotas get-service-quota --region us-east-1 --service-code ec2 --quota-code L-1216C47A --query 'Quota.Value'
Expected output
An error occurred (DryRunOperation) when calling the RunInstances operation: Request would have succeeded, but DryRun flag is set. # <- permission proven, nothing launched
640.0
The dry run checks the permission and launches nothing; the image id is one the account already runs. The quota is the most the account would launch before AWS stopped it: 640 vCPUs is 160 m5.xlarge instances. The hourly rate comes from the public on-demand pricing page on the day of the test, and the report shows the arithmetic, for example 160 instances at USD 0.192 an hour is about USD 22,400 a month, labelled as an estimate.
Negative result. UnauthorizedOperation on the dry run, or a quota close to what the account already uses.
Fix. Remove ec2:RunInstances from identities that do not launch compute, keep quotas near real use, and set AWS Budgets alerts on hourly spend.
Did we change only what we said we would?
aws cloudtrail lookup-events --region us-east-1 --start-time 2026-10-28T13:00Z \
--lookup-attributes AttributeKey=ReadOnly,AttributeValue=false \
| jq -r '.Events[] | select(.Username == "ci-deploy") | .CloudTrailEvent | fromjson
| [.eventTime, .eventName, (.errorCode // "-")] | @tsv'
Expected output
2026-10-28T14:40:12Z RunInstances Client.DryRunOperation
2026-10-28T14:33:02Z DeleteTags -
2026-10-28T14:31:40Z CreateTags - # <- the only writes: one tag, its removal and a dry run
This is the cloud version of the operator log check on the network impact page, and its output goes in the report appendix. Event history holds management events only, so the S3 read shows in data event logs if the client has them. The impact paragraph then follows the web impact page: what was shown first, what was estimated second.
Negative result. Any other write event under the identity needs an explanation before anything else is written.
Fix. For the client, not the tester: log S3 data events on sensitive buckets and keep management events in an organization trail or CloudTrail Lake, so this same query can be run over any identity, including an attacker's, beyond the 90 days event history keeps.
The defender sees
ListObjectsV2 and GetObject in S3 data event logs where they are on, with Decrypt in CloudTrail for the customer key. SimulatePrincipalPolicy under our review role. CreateTags and DeleteTags two minutes apart, and RunInstances with the error code Client.DryRunOperation, which should still alert, because a dry run from a key that never launches instances is reconnaissance. GuardDuty may raise a finding for API calls from an unusual location. If the client saw none of this, the missing alerts are a finding beside the impact. The CloudTrail and GuardDuty evidence review and the Entra sign-in and audit log review turn these events into the defender's view section of the report.
Done when
Read access is shown by one hashed object and a count, destructive permissions by a simulation, write access by a tag and its removal, cost by a dry run and a priced quota, and our own CloudTrail writes match that list exactly.
Not covered
Reaching the identity in the first place (see AWS IAM privilege escalation checks), destructive, ransomware or denial of wallet simulations, which this page rules out, and providers other than AWS and Azure; the outputs shown are representative examples on placeholder accounts, not captured runs.