Amazon GuardDuty is switched on in most AWS estates, yet many teams run it with gaps in Region coverage, protection plans left off, broad suppression rules and findings that nobody routes to a responder. This guide explains how GuardDuty is organised across accounts, what each protection plan watches, how to read finding names and severities, how to tune without hiding real attacks, and how to turn findings into a response, especially for the many findings that are really about stolen machine credentials. It is written for SOC analysts, cloud detection engineers and the platform teams who own the GuardDuty configuration. For the wider AWS control set, start with the AWS Security Guide; this page goes deep on GuardDuty alone.
This is an independent guide from NHI Mgmt Group, based on AWS's public documentation as of November 2026. It is not sponsored or reviewed by AWS. Product features change, so confirm settings against AWS's current documentation.
Key takeaways
- GuardDuty is Regional: it must be enabled, with the same delegated administrator account, in every Region you use and every Region you do not, because attackers launch resources where nobody is looking.
- Protection plans decide what you can see: foundational detection covers CloudTrail management events, VPC flow logs and DNS logs; S3 data events, EKS audit logs, runtime events, RDS logins, Lambda network activity and malware scanning each need their plan enabled.
- Many high-value findings are identity findings: credential exfiltration, anomalous API behaviour and root credential usage point at a compromised role, user or access key, so the response is usually credential revocation, not host cleanup.
- Suppression is not free: suppressed findings are archived, are not exported to EventBridge or Security Hub CSPM, and are not used by Extended Threat Detection to build attack sequences.
- Routing is part of the control: new findings reach EventBridge in near real time, but repeat occurrences are batched (default every 6 hours), so set the update frequency deliberately and send Critical and High findings straight to on-call.
How identity and access work in Amazon GuardDuty
GuardDuty is a managed detection service. Once enabled it reads the foundational data sources directly from AWS (you do not need to turn on CloudTrail trails or VPC flow logs yourself for GuardDuty to analyse them) and produces findings. Two identity questions matter: which account administers GuardDuty for the organisation, and which principals can change its configuration.
Account model
- Delegated administrator: the organisation management account designates a member account as the delegated GuardDuty administrator. AWS advises against using the management account itself. A dedicated security tooling account is the usual choice. See the AWS Multi-Account Security Reference Architecture Guide.
- Same account in every Region: GuardDuty is Regional while AWS Organizations is global, so the delegated administrator must be set in each Region, and AWS requires it to be the same account everywhere.
- Auto-enable: the administrator can auto-enable GuardDuty and individual plans for
ALLmembers (new and existing) orNEWmembers only.NEWhas an ordering trap in opt-in Regions: an account that joined the organisation before opting in to the Region is not "new" there and stays uncovered.ALLavoids that. - Scale: one delegated administrator can manage up to 50,000 member accounts.
- Administrator-only settings: in a multi-account setup only the administrator can create suppression rules and change the EventBridge update frequency, which keeps members from quietly hiding their own findings.
Foundational sources and protection plans
| Coverage | Data analysed | Typical findings | Enable when |
|---|---|---|---|
| Foundational threat detection | CloudTrail management events, VPC flow logs (EC2), DNS logs | Credential misuse, crypto-mining traffic, calls from known malicious IPs, root usage | Always, every account and Region |
| Extended Threat Detection | Correlates signals across sources, resources and time in an account | Attack sequence findings (Critical) | On automatically at no extra cost |
| S3 Protection | CloudTrail S3 data events | Exfiltration:S3/AnomalousBehavior, Impact:S3/AnomalousBehavior.Delete | Any account holding sensitive buckets |
| EKS Protection | EKS audit logs | Anonymous access, suspicious Kubernetes API use | Any account running EKS |
| Runtime Monitoring | Operating system events on EKS, EC2 and ECS (including Fargate) | Process, file and network behaviour such as reverse shells or miners | Production compute; not on by default |
| Malware Protection for EC2 | EBS volumes attached to instances | Malware on an instance that raised a finding, or on demand | EC2 workloads |
| Malware Protection for S3 | Newly uploaded objects | Malicious uploads to buckets that accept external files | Upload buckets; usable without GuardDuty itself |
| Malware Protection for AWS Backup | EBS snapshots, AMIs and recovery points | Malware in backups before restore | Backup and recovery accounts |
| RDS Protection | Login activity on supported Aurora and RDS databases | Anomalous or brute-force logins | Accounts with RDS or Aurora |
| Lambda Protection | Lambda network activity, starting with VPC flow logs | Functions mining or talking to malicious hosts | Serverless workloads |
| AI Protection | Use of Amazon Bedrock, Bedrock AgentCore and SageMaker AI | Anomalous model invocations, cost harvesting | Accounts with AI workloads |
When GuardDuty is first enabled, all plans except Runtime Monitoring are turned on and included in the 30-day free trial. After the trial they keep running and charging, which is when teams are tempted to switch plans off. Decide per plan on risk, not on the first bill: S3 Protection on a log archive account with no sensitive data may be low value, while on a data lake account it is essential.
Admin roles and privileged access
An attacker who can change GuardDuty can blind it. The actions below belong to the security team only and should be protected by SCPs in every member account.
| Action | Risk if misused | Guardrail |
|---|---|---|
guardduty:DeleteDetector, guardduty:UpdateDetector | Turns detection off or disables plans in an account | SCP deny for all except the security administration role |
guardduty:DisassociateFromAdministratorAccount, guardduty:DisassociateMembers | Removes an account from central visibility | SCP deny in member accounts |
guardduty:CreateFilter, guardduty:UpdateFilter | Suppression rules that archive real findings | Administrator account only; peer review; tag each rule with a reason |
guardduty:CreateIPSet, guardduty:UpdateIPSet | A trusted IP list stops findings for traffic from those addresses | Restrict; review entries quarterly |
guardduty:UpdatePublishingDestination, guardduty:DeletePublishingDestination | Breaks long-term export of findings | Restrict; alert on change |
guardduty:ArchiveFindings | Hides findings from the active view | Allow analysts, but log and review who archives what |
GuardDuty's own changes appear in CloudTrail, so alert on any of these calls made outside the security role. Removing or changing the delegated administrator does not disable GuardDuty in member accounts, but it does break central visibility until the new administrator is associated in every Region.
Integrations, API tokens and service accounts
GuardDuty touches identity in two directions: it uses AWS-managed access to your data sources, and a large share of its findings concern your machine identities.
Findings that are really about credentials
| Finding type | What it means | First response |
|---|---|---|
UnauthorizedAccess:IAMUser/InstanceCredentialExfiltration.OutsideAWS | Credentials issued to an EC2 instance role are being used from an IP outside AWS | Revoke the role's active sessions, investigate the instance, check IMDSv2 and SSRF exposure |
UnauthorizedAccess:IAMUser/InstanceCredentialExfiltration.InsideAWS | Instance role credentials used from a different AWS account | As above; check remoteAccountDetails.affiliated to see if the account is yours |
UnauthorizedAccess:IAMUser/ResourceCredentialExfiltration.OutsideAWS / .InsideAWS | Credentials from another AWS resource used from an unexpected location | Revoke sessions for the resource's role; investigate the resource |
CredentialAccess:IAMUser/AnomalousBehavior and the other AnomalousBehavior types | A principal calling APIs in a way unusual for it (discovery, persistence, privilege escalation, impact) | Confirm with the owner; if unconfirmed, disable keys and contain |
CredentialAccess:IAMUser/CompromisedCredentials | Credentials believed compromised | Rotate immediately; hunt for new users, keys and roles |
Policy:IAMUser/RootCredentialUsage | Root user credentials used | Confirm a break-glass event; otherwise treat as account compromise |
Stealth:IAMUser/CloudTrailLoggingDisabled | A trail was stopped or deleted | Restore logging; treat the actor as hostile until proven otherwise |
When GuardDuty flags IAM credentials being used from an unusual network, responding to the first credential finding means hunting for every identity and resource the credential created, such as new IAM users, Lambda function URLs and instances with termination protection set, not only rotating the key. See the Cloud Access Key Leak Playbook and the EC2 Instance Metadata Service Hardening Guide.
Integrations
- Amazon EventBridge: every finding is published as an event with
sourceaws.guarddutyanddetail-typeGuardDuty Finding. Rules in the administrator account receive member findings too; filter ondetail.accountIdordetail.severityto route. - AWS Security Hub CSPM: aggregates GuardDuty findings with Macie, Inspector and configuration checks. See the AWS Security Hub Guide.
- Amazon Detective: builds behaviour graphs from logs to investigate a finding's root cause.
- Export to S3: GuardDuty keeps findings for 90 days; configure a publishing destination to an S3 bucket in the log archive account for longer retention.
- SIEM and SOAR: the automation that consumes findings usually runs as a role with power to disable keys and isolate instances. Scope it tightly and protect it as a privileged machine identity, because compromising it is a fast route to sabotaging your own response.
Logging and detection
Reading a finding
Finding types follow the pattern ThreatPurpose:ResourceTypeAffected/ThreatFamilyName.DetectionMechanism!Artifact, for example Backdoor:EC2/C&CActivity.B!DNS. The prefix tells you the stage (Recon, UnauthorizedAccess, CredentialAccess, Exfiltration, Impact, Stealth, Policy and so on) and the resource type tells you which runbook applies.
| Severity | Range | Meaning (AWS) | Suggested routing |
|---|---|---|---|
| Critical | 9.0 to 10.0 | An attack sequence may be in progress or recently occurred | Page on-call immediately; major incident process |
| High | 7.0 to 8.9 | A resource is compromised and actively used for unauthorised purposes | Page on-call; contain within the hour |
| Medium | 4.0 to 6.9 | Suspicious activity that deviates from normal behaviour | SOC queue; confirm with resource owner same day |
| Low | 1.0 to 3.9 | Attempted activity that did not compromise the environment | Record; review for trends |
Update frequency and archiving
New findings are sent to EventBridge in near real time. Later occurrences of the same finding are aggregated and sent at an interval the administrator sets to 15 minutes, 1 hour or the default 6 hours. For a SOC, 15 minutes is usually the right choice, so that an ongoing exfiltration keeps updating the ticket. Manually archived findings still go to EventBridge; findings archived by a suppression rule do not.
Response playbook outline
- Triage: identify the resource type, principal and account from the finding; check whether the activity is a known test or tool.
- Credential findings: for roles, revoke active sessions (an inline deny on tokens issued before now); for IAM users, deactivate access keys; then search CloudTrail for every action the principal took and every identity it created.
- Compute findings: isolate the instance with a restrictive security group, snapshot volumes for forensics (Malware Protection for EC2 can scan them), and replace the instance rather than clean it.
- Data findings: check what was read or deleted using S3 data events, and whether versioning or Object Lock allows recovery. See the Amazon S3 Security Guide.
- Close out: record root cause, and add or narrow a suppression rule only if the finding was a confirmed, repeated false positive.
Metrics
- Coverage: percentage of account and Region pairs with GuardDuty enabled and associated with the administrator (target 100%, including unused Regions).
- Plan coverage: percentage of accounts with S3, EKS, Runtime Monitoring and RDS Protection enabled where those services are used.
- Response speed: median time from a High or Critical finding to containment.
- Suppression health: number of suppression rules, the share scoped to a whole finding type with no second criterion (target zero), and the date each was last reviewed.
- Credential findings: count per month of credential exfiltration and root usage findings, which should fall as IMDSv2 and short-lived credentials roll out.
Hardening baseline
| Area | Baseline |
|---|---|
| Administration | Delegated administrator in a security tooling account, the same account in every Region; never the management account |
| Coverage | Auto-enable set to ALL in every Region, including Regions you do not use (deny those Regions with an SCP as well) |
| Plans | S3, EKS, RDS, Lambda and Malware Protection where the services run; Runtime Monitoring on production compute with the agent managed automatically where possible |
| Tamper protection | SCP denying detector deletion, disassociation and filter or IP set changes outside the security role |
| Routing | EventBridge rules for severity 7 and above to paging; all findings to SIEM; update frequency 15 minutes |
| Retention | Publishing destination to an S3 bucket in the log archive account, encrypted with a KMS key that GuardDuty is allowed to use |
| Tuning | Suppression rules only for repeated, confirmed false positives, always pairing the finding type with a second criterion |
Rollout in phases
- Centralise: designate the delegated administrator in every Region and associate all existing accounts.
- Baseline for two weeks: enable foundational detection and the relevant plans with no suppression; let anomaly models learn and collect the false positive profile.
- Route: connect EventBridge to SIEM and paging; agree runbooks per resource type with named owners.
- Tune narrowly: add suppression rules for known scanners, bastions and on-premises egress addresses, each with a second criterion such as an AMI ID, tag or source IP.
- Lock: apply the tamper-protection SCP in a pilot OU, then across the organisation. Rollback is detaching the SCP.
- Exercise: generate sample findings in a test account and confirm each reaches the right queue and runbook.
Choosing coverage by organisation size
- Small estate: a single account can enable GuardDuty in every Region, send High and Critical findings by email or chat through EventBridge and SNS, and still get value from foundational detection and S3 Protection.
- Large estate: the cost of S3 Protection and Runtime Monitoring scales with data events and compute, so decide by account class (production data, production compute, sandbox) and record exceptions. Sandbox accounts still need foundational detection: they are often where stolen keys are first tested.
Framework mapping
| Activity | NIST CSF 2.0 | NIST SP 800-53 Rev. 5 | ATT&CK examples |
|---|---|---|---|
| Continuous monitoring of cloud activity | DE.CM-01, DE.CM-09 | SI-4 | T1078.004 Cloud Accounts, T1580 Cloud Infrastructure Discovery |
| Correlation and analysis | DE.AE-02, DE.AE-03 | AU-6, IR-5 | T1552.005 Cloud Instance Metadata API |
| Malware scanning | DE.CM-09 | SI-3 | T1496.001 Compute Hijacking |
| Triage and response | RS.MA-02, RS.AN-03 | IR-4 | T1530 Data from Cloud Storage, T1485 Data Destruction |
In CIS Controls v8.1 GuardDuty supports Control 13 (Network Monitoring and Defense) and Control 8 (Audit Log Management); in ISO/IEC 27001:2022 it supports A.8.16 Monitoring activities.
Common misconfigurations
- Only enabling home Regions: attackers with stolen keys often launch miners in Regions the victim never uses, where no detector is running.
- Using
NEWauto-enable with opt-in Regions: accounts that joined before the Region was opted in remain uncovered without anyone noticing. - Whole-type suppression: a rule that archives every
Recon:EC2/Portscanor every EKS finding hides real attacks and removes those signals from attack sequence correlation. - Findings with no destination: findings sit in the console while nobody is paged; the 90-day retention then quietly discards the evidence.
- Treating credential findings as host problems: rebuilding the instance while the exfiltrated role session stays valid leaves the attacker in place.
- Disabling plans after the trial without a risk decision, leaving S3 data access and runtime behaviour invisible in the accounts that need them most.
Practitioner checklist
- Designate one delegated administrator account, not the management account, in every Region.
- Set auto-enable to
ALLand verify withListMembersthat every account is associated in every Region. - Enable the protection plans that match the services each account runs, including Runtime Monitoring on production compute.
- Deploy an SCP that denies detector deletion, disassociation and filter or IP set changes outside the security role.
- Route severity 7 and above to paging and all findings to the SIEM; set the update frequency to 15 minutes.
- Export findings to an encrypted S3 bucket in the log archive account for retention beyond 90 days.
- Write runbooks for credential exfiltration, anomalous IAM behaviour, root usage and S3 impact findings.
- Allow suppression rules only for confirmed, repeated false positives with a second criterion and a review date.
- Test the pipeline quarterly with sample findings and record time to containment.
Standards and references
- AWS: What is Amazon GuardDuty?
- AWS: Managing GuardDuty accounts with AWS Organizations
- AWS: Severity levels of GuardDuty findings
- AWS: GuardDuty IAM finding types
- AWS: Suppression rules in GuardDuty
- AWS: Processing GuardDuty findings with Amazon EventBridge
- AWS: GuardDuty S3 Protection
- MITRE ATT&CK: Cloud Accounts
Related NHI Mgmt Group resources: AWS Security Guide · Cloud Detection Engineering Guide · AWS CloudTrail Guide · Detection Engineering for Non-Human Identities Guide