
Most attacks on an Amazon Web Services (AWS) account do not look like malware. They look like ordinary API calls made with stolen keys: a new user here, a changed policy there, a trail switched off so nobody sees the rest. AWS CloudTrail records almost every one of those calls. The problem is that a busy account writes millions of them a day, so the record is only useful once something reads it and reacts.
That reading is the job of a SIEM (security information and event management) tool. This guide lists 18 AWS CloudTrail SIEM use cases worth running, the CloudTrail event each one reads, and what the detection is really looking for. For the on-premises side of the same account, see our guides to Linux SIEM use cases and Active Directory SIEM use cases.
The short answer: the AWS CloudTrail detections that catch the most real attacks are someone turning off CloudTrail, Config or GuardDuty; root account use, especially without multi-factor authentication; a console login from an impossible location; new IAM users, access keys or policy attachments; a trust policy rewritten to let an outside account assume a role; an S3 bucket made public; and bulk reads or snapshot exports that move data out. Almost all of these read CloudTrail management events, which AWS records by default. The exceptions are bulk object reads, which need Amazon S3 data events, and GuardDuty findings, which need GuardDuty switched on.
Where each detection catches an attacker in AWS
The CloudTrail use cases along a typical account takeover, with the source each one reads
- Get inUse stolen keys or a passwordRoot login
ConsoleLoginRoot use without MFAmanagement eventsImpossible travelConsoleLogin + geoAssume-role brute forceUpdateAssumeRolePolicy - Look aroundMap the accountBurst of Describe and List calls
management eventsGuardDuty high-severity findingGuardDuty - Gain privilegeBecome adminNew IAM user, key or policy
CreateUser, CreateAccessKeyPolicy or role attachmentAttachUserPolicyTrust policy rewrittenUpdateAssumeRolePolicySSO identity provider changesso.amazonaws.com - StayKeep a way back inLogin profile added
CreateLoginProfileLambda function URL backdoorCreateFunctionUrlConfigRoute 53 domain transferroute53domains - Blind youStop the recordingCloudTrail stopped or deleted
StopLogging, DeleteTrailConfig or GuardDuty disabledmanagement eventsFlow log or log group deletedmanagement events - Take dataExpose or copy it outS3 bucket made public
PutBucketAclBulk object readsS3 data eventsSnapshot shared or exportedModifySnapshotAttribute
Before you start: the AWS logs these detections need
A detection cannot fire on an event that was never recorded. In AWS that is the usual reason a cloud use case fails, so start with the sources.
| Source | What it records | On by default? |
|---|---|---|
| CloudTrail management events | Control-plane API calls: logins, IAM changes, resource creation and deletion, config changes | Yes. The first copy delivered to a trail is free |
| CloudTrail data events | Data-plane actions inside a resource, such as S3 object reads and writes or Lambda invocations | No. You turn them on per resource, and they are charged |
| CloudTrail Event history | A searchable record of management events for the last 90 days, per region | Yes, and free, but it is not a feed a SIEM can read continuously |
| Amazon GuardDuty | Findings built from CloudTrail, VPC Flow Logs and DNS query logs, scored by severity | No. You enable it, then route findings out |
| VPC Flow Logs | Source, destination, port and byte counts for network traffic in a virtual private cloud | No. You create them and send them to CloudWatch Logs or S3 |
To read any of this continuously, create a CloudTrail trail that delivers to an Amazon S3 bucket or to CloudWatch Logs, then forward that to your SIEM. Event history alone is not enough, because it only answers searches and keeps 90 days. The figure below shows the path.
How CloudTrail reaches a SIEM
Event history is for ad hoc searches. A trail gives a SIEM a continuous feed
- API callSomeone acts in the account
- CloudTrailRecords the event
- TrailDelivers to S3 or CloudWatch Logs
- SIEM reads the feedParses and normalizes
- Detection firesRule matches, alert raised
The 18 use cases
Each row is a detection, the CloudTrail event or source it reads, and what it catches. Every one of these ships as a built-in rule in UTMStack, mapped to the technique in the MITRE ATT&CK framework shown. ATT&CK is a public catalog of attacker techniques.
Blind the defender first
An attacker who has admin rights often starts by turning off the recording. These are the highest-value detections, because they protect everything else.
| Use case | CloudTrail event | What it catches | ATT&CK |
|---|---|---|---|
| CloudTrail logging disabled | StopLogging, DeleteTrail | Someone stops or deletes the trail so later actions are not recorded | T1562.008 |
| Event selector changed | PutEventSelectors | A trail quietly narrowed so it stops recording the events that matter | T1562.008 |
| Config recorder stopped | StopConfigurationRecorder | AWS Config, which tracks resource changes, switched off | T1562.001 |
| GuardDuty detector deleted | DeleteDetector | The threat-detection service removed so its findings stop | T1562.008 |
| Flow log deleted | DeleteFlowLogs | Network logging for a virtual private cloud removed | T1562.008 |
| CloudWatch log group deleted | DeleteLogGroup | Stored logs destroyed to cover tracks | T1070 |
| WAF rule or ACL deleted | DeleteWebACL, DeleteRule | A web application firewall rule removed to open an app back up | T1562.001 |
Get in and move up
| Use case | CloudTrail event | What it catches | ATT&CK |
|---|---|---|---|
| Root account used without MFA | management events, userIdentity type Root | The most powerful account in use, with no second factor | T1078.004 |
| Console login impossible travel | ConsoleLogin from two countries within 30 minutes | The same user signing in from places too far apart to be real | T1078.004 |
| Assume-role policy brute force | UpdateAssumeRolePolicy, 5 in 15 minutes | Repeated edits to a role’s trust policy, probing for a way in | T1110 |
| IAM backdoor creation | CreateUser, CreateAccessKey, AttachUserPolicy, 3 in 30 minutes | A fresh user with keys and a policy, built to keep access | T1136.003 |
| IAM privilege escalation path | CreatePolicyVersion, AttachRolePolicy, AddUserToGroup | Permissions widened step by step toward full admin | T1098 |
| SSO identity provider change | sso.amazonaws.com directory changes | The single sign-on trust changed, a route to a golden SAML attack | T1078.004 |
Reconnaissance, persistence and data theft
| Use case | CloudTrail event | What it catches | ATT&CK |
|---|---|---|---|
| Burst of Describe and List calls | 50 read-only calls in 10 minutes | An attacker mapping the account right after getting in | T1106 |
| Lambda function URL backdoor | CreateFunctionUrlConfig | A public web address bolted onto a function for later access | T1546 |
| S3 bucket made public | PutBucketAcl, DeleteBucketPublicAccessBlock | A bucket opened to the internet, exposing its contents | T1530 |
| S3 bulk data exfiltration | GetObject, 100 in 15 minutes (data events) | Large-scale copying of objects out of a bucket | T1530 |
| Snapshot shared or exported | ModifySnapshotAttribute, EC2 and RDS export | A disk or database copy shared to an outside account or exported | T1537 |
How UTMStack fits, and where it stops
UTMStack ships 73 AWS detection rules, each mapped to MITRE ATT&CK and checked the moment a CloudTrail event arrives, before it is written to storage. The 18 above are the ones worth turning on first. You ingest AWS by creating a trail and letting UTMStack read it through CloudWatch, the same way its threat detection reads any other source. GuardDuty findings and VPC Flow Logs come in on the same feed.
Three detections in the list deserve plain warnings:
- Bulk object reads need S3 data events. The S3 exfiltration rule counts GetObject calls, and GetObject is a data event. AWS does not record data events by default, and they are charged. If you have not turned them on for the bucket, the rule has nothing to read. Management-event detections, which are most of the list, work without this.
- Impossible travel needs location data. The rule compares the country of two logins. UTMStack adds the location from the source address, so a login through a virtual private network can look like travel. Treat it as a prompt to check, not proof.
- Snapshot and packet-capture detections depend on the right events reaching the trail. Cross-account snapshot sharing shows up in management events, but full network packet capture and some export paths rely on sources you may not be sending yet.
Being honest about limits matters more in the cloud, where it is easy to assume a managed service is watching when it is not. GuardDuty, for example, only produces findings for the data sources it is given, and only when it is enabled in every region you use.
What UTMStack does once a rule fires is the second half of the job. Its automated response flows can disable an IAM user or revoke a session so a stolen key stops working while someone investigates. For how the matching works, see our guide to SIEM correlation rules, and for the wider picture of which logs to collect, what logs should be sent to a SIEM. UTMStack is priced per device, never by log volume, so a noisy AWS account does not change the bill; for a quote, see pricing.
Frequently asked questions
Does AWS CloudTrail count as a SIEM? No. CloudTrail records events and keeps 90 days of management events in Event history that you can search. It does not correlate events across sources, raise alerts on patterns, or respond. A SIEM reads the CloudTrail feed and does that work. GuardDuty is closer to detection, but it covers only its own data sources and does not take in your firewall, server or identity logs.
What is the difference between CloudTrail management events and data events? Management events are control-plane actions, such as creating a user or changing a policy, and the first copy to a trail is free. Data events are actions inside a resource, such as reading an S3 object or invoking a function. Data events are off by default and are charged, so you turn them on only for the buckets and functions that hold sensitive data.
Which AWS logs should I send to a SIEM first? Start with CloudTrail management events from every region, because they carry the logins, IAM changes and service-disabling actions in most of these use cases. Add GuardDuty findings next, then VPC Flow Logs, then data events for your most sensitive S3 buckets. Sending every data event from the start is expensive and rarely worth it.
Can a SIEM stop an AWS attack, not just alert on it? It can respond if it is allowed to act in the account. UTMStack response flows can disable an IAM user or revoke a session so stolen credentials stop working. Response should be scoped carefully, because a wrong automatic action in a production account can cause an outage.
How far back can I investigate an AWS incident? Event history keeps 90 days of management events for searching. For anything longer, you need a trail delivering to storage or a SIEM that retains the data. UTMStack keeps events for up to five years with searchable cold storage, which covers most compliance and investigation needs.
Sources
- AWS, “What Is AWS CloudTrail?” and CloudTrail concepts: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-user-guide.html
- AWS, “Logging management events” and “Logging data events” for CloudTrail: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/logging-management-events-with-cloudtrail.html
- AWS, “Working with CloudTrail Event history”: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/view-cloudtrail-events.html
- Amazon GuardDuty, “Foundational data sources”: https://docs.aws.amazon.com/guardduty/latest/ug/guardduty_data-sources.html
- MITRE ATT&CK, Enterprise techniques: https://attack.mitre.org/techniques/enterprise/


