study guide
AWS Certified Security – Specialty · SCS-C03 · current exam since Dec 2, 2025

SCS-C03 Study Guide

For the AWS (Amazon Web Services) Certified Security – Specialty exam. Every task and skill from the current official SCS-C03 exam guide, explained simply, with the trade-offs the questions test. Pick a domain tab or search any service. Yellow boxes are memory hooks; dashed boxes are exam tips; orange boxes flag changes AWS made after the guide was written.

No matches. Try a shorter word, a service name, or an acronym.
Start here

How the exam works

Format, scoring, what changed from SCS-C02, what's in scope, and which services AWS has changed since the guide was published.

The exam at a glance

A specialty-level exam for people who secure AWS workloads. Questions are scenarios: a company has a security requirement plus constraints (least effort, lowest cost, no code changes, across every account). Often two answers would work; pick the one that meets every constraint. The guide targets the equivalent of 3–5 years of experience securing cloud solutions.

65questions (50 scored + 15 unscored)
170minutes, so about 2.6 minutes per question
750passing scaled score (100–1,000)
0penalty for guessing; blanks count as wrong
$300USD (50% off voucher if you hold an active AWS cert)
4question types, two of them new

Question types

TypeWhat you do
Multiple choicePick 1 correct answer out of 4.
Multiple responsePick 2 or more correct answers out of 5 or more. All must be right.
Ordering New in C03Pick the correct 3–5 steps and put them in the right order (for example, the steps to contain a compromised instance). All-or-nothing.
Matching New in C03Match 3–7 prompts to responses (for example, threats to the service that stops each). Every pair must be right.

Scoring is compensatory: you only need to pass overall, not each domain. The 15 unscored questions aren't marked, so treat every question as real. Delivered at a Pearson VUE test center or online proctored.

The domain tabs include ordering and matching practice widgets, because those formats punish half-knowledge: one step out of place scores zero. (The certification web page still says "multiple choice or multiple response"; the official exam guide lists all four types, so prepare for all four.)
Read the last sentence first ("MOST secure", "LEAST operational overhead", "with the FEWEST changes"). Then read the scenario hunting for the constraints that kill wrong answers.
The Simplified Chinese, Spanish (Latin America) and Portuguese (Brazil) versions of this exam retire after December 31, 2026. English, Japanese and Korean continue.

Where the points are

DomainWeightApprox. scored questions
1. Detection16%~8
2. Incident Response14%~7
3. Infrastructure Security18%~9
4. Identity and Access Management20%~10
5. Data Protection18%~9
6. Security Foundations and Governance14%~7

IAM is now the single biggest domain (up from 16% to 20%). Policy evaluation logic, cross-account access and troubleshooting "access denied" are worth the most study time.

How SCS-C02 domains map to SCS-C03

SCS-C02 domain%SCS-C03 domain%
1. Threat Detection and Incident Response14%1. Detection16%
2. Security Logging and Monitoring18%2. Incident Response14%
3. Infrastructure Security20%3. Infrastructure Security18%
4. Identity and Access Management16%4. Identity and Access Management20%
5. Data Protection18%5. Data Protection18%
6. Management and Security Governance14%6. Security Foundations and Governance14%

The old "threat detection + IR" and "logging + monitoring" domains were split by job: everything about seeing problems is now Domain 1 (Detection), everything about acting on them is Domain 2 (Incident Response).

What's new and what's gone

Added skills New in C03

SkillWhat it addsWhere
2.2.3Validate findings from AWS security services to assess scope and impactDomain 2
3.1.4Edge integrations: ingest data in OCSF format, third-party AWS WAF rulesDomain 3
3.2.7Guardrails for generative AI apps (OWASP Top 10 for LLM Applications)Domain 3
5.1.3Inter-resource encryption in transit (EMR, EKS, SageMaker AI, Nitro)Domain 5
5.3.3Imported key material vs AWS-generated key materialDomain 5
5.3.4Masking sensitive data (CloudWatch Logs data protection, SNS message data protection)Domain 5
5.3.5Keys and certificates across one or many Regions (multi-Region KMS keys, Private CA)Domain 5

Removed from the blueprint

  • AWS Security Finding Format (ASFF) and the AWS Security Incident Response Guide as named items
  • CloudTrail log format and components
  • Host-based security (host firewalls, OS hardening as its own topic)
  • Reachability Analyzer, TCP/IP fundamentals (TCP vs UDP, OSI model), network connectivity troubleshooting
  • Policy element basics (Principal, Action, Resource, Condition) as a stand-alone skill
  • TLS concepts, cross-Region networking with private and public VIFs, S3 static website hosting

"Removed" means not tested as its own skill. You still need policy elements to read a policy, and TLS basics to pick a load balancer security policy.

Services AWS changed after the guide was written

Several services named in the SCS-C03 guide have since been retired or closed to new customers. The exam can still use them as written (exam content lags product changes), so know what they do and what replaced them. In this guide they carry a Maintenance or Retired tag.
Service or feature in the guideWhat changedCurrent replacement
Amazon CodeGuru Security (skill 3.2.6)Retired Support ended November 20, 2025Amazon Inspector code security (SAST, dependency (SCA) and IaC scanning of GitHub/GitLab repos) and Amazon Q Developer security scans
AWS Audit Manager (skill 6.3.2)Maintenance No new customers from April 30, 2026AWS Config conformance packs (AWS's own recommendation), Security Hub CSPM standards; AWS Artifact for AWS's reports
AWS CloudTrail Lake (in-scope list)Maintenance No new customers from April 30, 2026Amazon CloudWatch (AWS's recommended migration; OCSF/OTel support), or Athena over the trail's S3 bucket, or Security Lake
SNS message data protection (skill 5.3.4)Maintenance No new customers from April 30, 2026CloudWatch Logs data protection policies for logs; mask in the publisher/subscriber (Lambda), Macie for data at rest
S3 Glacier Vault Lock (skill 5.2.2)Maintenance The original Amazon Glacier vault service closed to new customers November 7, 2025S3 Object Lock on S3 Glacier storage classes (fully available), AWS Backup Vault Lock
Systems Manager Incident ManagerMaintenance Closed to new customers November 7, 2025Systems Manager OpsCenter (named in the guide), EventBridge + SSM Automation runbooks
Shield Advanced automatic application-layer mitigationSuperseded from March 26, 2026 (legacy stays available to existing customers)AWS WAF Anti-DDoS managed rule group, which mitigates HTTP floods within seconds
If an answer uses a maintenance-mode service and another answer does the same job with a current service, the exam was almost certainly written with the guide's service in mind. Answer the question as asked; don't overthink product lifecycle.

In-scope and out-of-scope services

AWS's note on the list matters: security affects all AWS services. Many services aren't listed because the service itself is out of scope, but its security features are in scope. You won't be asked how to set up S3 replication, but you might be asked about an S3 bucket policy.

CategoryIn scope (SCS-C03)
AnalyticsAthena, OpenSearch Service
Application integrationSNS, Step Functions
ComputeAPI Gateway, EC2 (including EC2 Image Builder and EC2 Instance Connect), EKS, EMR, Lambda, Data Lifecycle Manager
Developer toolsAWS Fault Injection Service
Internet of ThingsAWS IoT Core
Machine learningAmazon Bedrock, CodeGuru Security, Amazon Q Business, Amazon Q Developer, SageMaker AI
Management & governanceCloudFormation, CloudTrail, CloudTrail Lake, CloudWatch, Config, Control Tower, Managed Grafana, Organizations, Resilience Hub, RAM, Service Catalog, Systems Manager, Trusted Advisor, User Notifications, Well-Architected Tool
Networking & content deliveryApplication Recovery Controller, VPC (Network Access Analyzer, network ACLs, security groups, VPC endpoints, Site-to-Site VPN, Flow Logs, Verified Access), Client VPN, CloudFront, Verified Permissions, Route 53 (including Resolver DNS Firewall), Direct Connect, Elastic Load Balancing, Transit Gateway
Security, identity & complianceArtifact, Audit Manager, ACM, CloudHSM, Cognito, Detective, Directory Service, Firewall Manager, Automated Forensics Orchestrator for Amazon EC2, GuardDuty, IAM, IAM Identity Center, Inspector, KMS, Macie, Network Firewall, Private CA, Secrets Manager, Security Hub, Security Lake, Shield, Shield Advanced, STS, WAF
Storage & data managementS3, AWS Backup, DataSync, EFS (including lifecycle policies), FSx for Lustre

Out of scope

Only two services: Amazon MWAA (Managed Workflows for Apache Airflow) and AWS Payment Cryptography. If either appears as an option, it's almost certainly a distractor.

Technologies and concepts that may appear

AWS CLI, AWS SDKs, AWS Management Console, secure remote access, certificate management, infrastructure as code (IaC).

Job tasks out of scope

Designing cryptographic algorithms, packet-level traffic analysis, architecting overall cloud deployments, managing end-user compute (WorkSpaces etc.), training machine learning models.

The classic answer-killers

Phrase in the questionWhat it usually means
"Least operational overhead" / "most efficient"Managed, native features over custom Lambda code, scripts, EC2-based agents or third-party tools
"Prevent" vs "detect"Prevent = SCP, RCP, declarative policy, permissions boundary, CloudFormation Hook, Block Public Access. Detect = Config rule, GuardDuty, Security Hub, Access Analyzer. Don't pick a detective control when they asked to prevent
"Across all accounts" / "including new accounts"Organizations + delegated administrator + auto-enable, StackSets with auto-deployment, Firewall Manager, Control Tower controls
"Immutable" / "cannot be deleted, even by root"S3 Object Lock compliance mode, Backup Vault Lock compliance mode, logically air-gapped vault
"Without changing application code"Server-side encryption, load balancer TLS termination, WAF, Secrets Manager managed rotation, network controls; not client-side encryption or SDK changes
"Preserve evidence" / "forensics"Isolate, don't terminate; snapshot and capture memory before stopping; deactivate (don't delete) keys
"Near real time"EventBridge rules on findings/API events, CloudWatch metric filters and alarms, not daily reports or scheduled scans
"Full control of keys" / "AWS must not access keys"CloudHSM or external key store (XKS); for "control of key material lifecycle" think imported key material
Security answers almost always layer: one preventive control plus one detective control plus automation. When a multiple-response question asks for two or three, look for that mix.
Sources
Domain 1

Detection

16% of scored content · Tasks 1.1–1.3: monitoring and alerting, logging, troubleshooting

1.1The detection services at a glance

Each service answers one question. Exam questions usually describe the question and expect you to name the service.

ServiceAnswersLooks atNeeds you to enable logs?
Amazon GuardDuty"Is someone attacking or misusing my account right now?"CloudTrail management events, VPC Flow Logs, DNS query logs (foundational) + optional protection plansNo. GuardDuty reads its own independent copy of the foundational sources
Amazon Inspector"What software vulnerabilities and network exposure do I have?"EC2, ECR images, Lambda functions, code repositoriesNo (uses SSM Agent or agentless snapshot scanning for EC2)
Amazon Macie"Where is sensitive data (PII, credentials) in S3, and is the bucket exposed?"S3 objects and bucket settingsNo
AWS Security Hub CSPM"Am I following best practices and compliance standards?"Controls backed by AWS Config rules; aggregates findings from other servicesYes: AWS Config must be recording
AWS Security Hub"Which risks matter most, and how do they connect?"Correlates findings (GuardDuty, Inspector, Macie, CSPM) into exposures and attack paths in OCSFDepends on the services feeding it
Amazon Detective"What happened, how did it start, what else did it touch?"Behavior graph built from CloudTrail, VPC Flow Logs, GuardDuty findings, EKS audit logsNo; requires GuardDuty enabled
Amazon Security Lake"Where do I keep and query all security data in one standard format?"Normalizes AWS and third-party logs to OCSF in S3It collects the sources for you
AWS Config"What changed, who changed it, and is it compliant?"Resource configuration history and rulesConfig recorder must be on
IAM Access Analyzer"Who outside (or inside) my trust zone can reach this resource? What access is unused?"Resource policies, IAM policies, last-accessed dataNo
GuardDuty = the burglar alarm. Inspector = the building inspector (weak locks). Macie = the sniffer dog for sensitive data in S3. Security Hub = the control room. Detective = the investigator who replays the footage. Security Lake = the evidence archive.
Try it

Pick the detection service

1.1Amazon GuardDuty in depth

ProtectionWhat it adds
Foundational threat detectionAlways on once enabled: CloudTrail management events, VPC Flow Logs, Route 53 Resolver DNS logs. Detects reconnaissance, crypto-mining DNS lookups, credential use from unusual places, calls from Tor or known-malicious IPs.
S3 ProtectionWatches S3 data-plane events (GetObject, ListBucket…) for exfiltration patterns and suspicious policy changes
EKS ProtectionAnalyzes EKS audit logs (control plane) for suspicious Kubernetes API activity
Runtime MonitoringLightweight security agent on EKS, ECS (including Fargate) and EC2 that sees process, file and network activity inside the workload
Malware Protection for EC2Agentless scan of EBS volumes when GuardDuty sees suspicious behavior (or on demand)
Malware Protection for S3Scans newly uploaded objects in chosen buckets; can tag objects so a bucket policy blocks infected files
RDS ProtectionAnomalous login activity on Aurora / RDS
Lambda ProtectionNetwork activity from Lambda functions (for example, talking to crypto-mining pools)
Extended Threat DetectionOn automatically. Correlates signals over time into attack sequence findings (critical severity), for example "privilege discovery on S3, then bucket policy loosened, then data read from an unusual location". More plans enabled = more sequences it can see.
  • Finding type format: ThreatPurpose:ResourceType/ThreatFamily.Variant, for example CryptoCurrency:EC2/BitcoinTool.B!DNS or UnauthorizedAccess:IAMUser/InstanceCredentialExfiltration.OutsideAWS.
  • Tuning: trusted IP lists (no findings for these IPs), threat lists (your known-bad IPs), suppression rules (auto-archive known benign findings; they're still generated, just archived).
  • Multi-account: designate a delegated administrator (security tooling account) from the management account, then auto-enable for all existing and new member accounts. GuardDuty is Regional: enable it in every Region, including ones you don't use.
  • Acting on findings: every finding is sent to EventBridge automatically → rule → SNS, Lambda, Step Functions or SSM Automation.
"Detect EC2 instances talking to crypto-mining pools with the least effort" → enable GuardDuty. You do not need to turn on VPC Flow Logs or DNS logging first; that's a classic distractor.

1.1Security Hub, Macie, and building alerts

Security Hub CSPM vs Security Hub

Security Hub CSPMSecurity Hub
JobRuns best-practice controls against standards: AWS Foundational Security Best Practices, CIS AWS Foundations, PCI DSS, NIST SP 800-53Correlates findings across services into prioritized exposures and attack paths; response workflows
Finding formatASFF (AWS Security Finding Format)OCSF (Open Cybersecurity Schema Framework)
PrerequisiteAWS Config recording the relevant resource typesFeeds from GuardDuty, Inspector, Macie, CSPM
  • Central configuration from the delegated admin: one policy enables standards and controls across accounts, OUs and Regions.
  • Cross-Region aggregation: link Regions to one home Region to see everything in one place.
  • Automation rules: change finding fields automatically (suppress, set severity, add notes) when criteria match.
  • Custom actions send a chosen finding to EventBridge so an analyst can click "Isolate instance" and trigger a runbook.

Amazon Macie

  • Automated sensitive data discovery samples objects across all buckets continuously and gives each bucket a sensitivity score (cheap, broad).
  • Sensitive data discovery jobs scan chosen buckets fully, once or on a schedule (deep, targeted).
  • Managed data identifiers (credit cards, passport numbers, AWS secret keys…), custom data identifiers (your regex plus keywords), allow lists (ignore known test values).
  • Two finding kinds: policy findings (bucket made public, encryption removed) and sensitive data findings. Results go to EventBridge and Security Hub; full results to an S3 bucket you choose (KMS-encrypted).

Alerts and dashboards

NeedBuild it with
Alert within minutes when the root user signs in or someone calls an API without permissionCloudTrail → CloudWatch Logs → metric filter → CloudWatch alarm → SNS (the classic CIS benchmark alarms), or an EventBridge rule on the CloudTrail event
React to each GuardDuty / Security Hub / Macie findingEventBridge rule (filter on severity, finding type) → SNS, Lambda, Step Functions
Catch unusual spikes without a fixed thresholdCloudWatch anomaly detection alarms; CloudWatch Logs anomaly detection
Workload health monitoring (skill 1.1.2)Route 53 health checks, ELB target health checks, CloudWatch Synthetics canaries, alarms on 4xx/5xx spikes
One dashboard across accountsCloudWatch cross-account observability (monitoring account), Amazon Managed Grafana
Route notifications to people centrallyAWS User Notifications (email, chat, mobile) from EventBridge events

1.1Automated assessments and investigations

ToolWhat it automatesExample
AWS Config rulesContinuous or periodic evaluation of resource configuration. Managed rules, custom Lambda rules, or custom rules written in Guard (policy-as-code)s3-bucket-public-read-prohibited, restricted-ssh, encrypted-volumes
Conformance packsA bundle of Config rules + remediation deployed as one unit, to one account or the whole organizationOperational Best Practices for PCI DSS / HIPAA / NIST sample packs
Config aggregatorMulti-account, multi-Region view of config and complianceOrg-wide compliance dashboard in the security account
Security Hub CSPMScheduled and change-triggered control checks with a security scoreWeekly score report to leadership
Systems Manager State ManagerKeeps instances in a defined state on a schedule (associations)Make sure the CloudWatch agent and EDR agent are installed and running on every instance; re-apply if removed
Amazon InspectorContinuous rescans when new CVEs are published or new software is installedNo scheduled scans to manage
Trusted AdvisorSecurity checks (open ports, MFA on root, exposed keys)Org-wide view in the management account
"Make sure every instance always runs the security agent, even if someone uninstalls it" → State Manager association. "Evaluate every account against PCI DSS and remediate" → conformance pack (organization-wide) or Security Hub CSPM PCI standard + automation.

1.2Log sources: what each one captures

LogCapturesWatch out for
CloudTrail management eventsControl-plane API calls: who, what, when, from where (CreateUser, PutBucketPolicy, RunInstances)On by default (90-day Event history). Create a trail to keep them longer in S3
CloudTrail data eventsData-plane calls: S3 GetObject/PutObject, Lambda Invoke, DynamoDB item operations…Off by default, extra cost; use advanced event selectors to limit to sensitive buckets
CloudTrail Insights eventsUnusual API call volume or error ratesOff by default
CloudTrail network activity eventsAPI calls made through VPC endpoints, including denied calls from outside your orgHelps prove data-perimeter enforcement
VPC Flow LogsIP traffic metadata: source/dest IP and port, protocol, bytes, ACCEPT/REJECT (per VPC, subnet or ENI)No packet contents. Not captured: traffic to the Amazon DNS server, instance metadata (169.254.169.254), Amazon Time Sync, DHCP
Transit gateway flow logsTraffic through a transit gateway between VPCs and on-premisesThe east-west view between VPCs
Route 53 Resolver query logsDNS queries made from your VPCsSpot DNS exfiltration and calls to malicious domains
DNS Firewall logsQueries blocked or allowed by Resolver DNS FirewallPart of Resolver query logging
ELB access logsEvery request to an ALB/NLB/CLB: client IP, latency, path, TLS cipherOff by default; bucket policy must allow the ELB log delivery service
CloudFront standard / real-time logsEvery viewer request at the edgeReal-time logs go to Kinesis Data Streams
AWS WAF logsEvery inspected request, matched rule, actionDestination name must start with aws-waf-logs- (CloudWatch log group, S3 bucket or Firehose stream)
API Gateway execution / access logsExecution: request processing details. Access: who called what (custom format)Needs an account-level IAM role ARN set in API Gateway settings
S3 server access logsRequests to a bucket, best effort, no cost except storageCloudTrail data events are more complete and timely for security
EKS control plane logsAPI server, audit, authenticator, controller manager, schedulerOff by default; audit + authenticator are the security ones
Network Firewall logsAlert, flow and TLS logsAlert logs show which stateful rule fired
OS / application logsAnything on the instanceShip with the CloudWatch agent (needs an IAM role and config file)
Try it

Which log would show it?

1.2Organization trails and a central logging account

Management (or delegated admin) account creates an organization trail→Every member account, every Region→S3 bucket in the Log Archive account→Optional: CloudWatch Logs for alarms
  • Member accounts can see the org trail but can't modify or delete it.
  • Bucket policy grants cloudtrail.amazonaws.com s3:PutObject, scoped with aws:SourceArn / aws:SourceAccount (stops the confused deputy problem).
  • Encrypt with SSE-KMS: the KMS key policy must let CloudTrail GenerateDataKey* and let readers Decrypt. A missing key-policy statement is the #1 cause of "trail stopped delivering".
  • Turn on log file integrity validation: CloudTrail writes hourly digest files signed with SHA-256 so you can prove logs weren't changed or deleted (aws cloudtrail validate-logs).
  • Protect the bucket: separate account, restrictive bucket policy, versioning, S3 Object Lock, lifecycle to Glacier storage classes.

Centralizing CloudWatch Logs

  • Subscription filters stream log events in near real time to Lambda, Kinesis Data Streams, Firehose or OpenSearch. For cross-account, the central account creates a destination (with a resource policy allowing source accounts), and each source account subscribes its log groups to it.
  • CloudWatch agent: install with Systems Manager (Distributor or Run Command), store the agent config in Parameter Store, attach CloudWatchAgentServerPolicy to the instance role. In private subnets add a logs interface VPC endpoint (and monitoring for metrics).
  • Set a retention period on every log group (the default is never expire), and encrypt log groups with KMS where required.
Logs go out of the account that made them and into an account the workload team can't touch. Think "flight recorder kept in a different plane."

1.2Amazon Security Lake and third-party tools

Security Lake builds a security data lake in your own account: it collects logs, converts them to OCSF in Apache Parquet, and stores them in S3 managed by Lake Formation.

  • Native sources: CloudTrail management and data events, VPC Flow Logs, Route 53 Resolver query logs, Security Hub findings, EKS audit logs, AWS WAF logs.
  • Custom sources: third-party or on-premises data, already converted to OCSF.
  • Rollup Regions consolidate data from many Regions; set retention and storage-class transitions.
  • Subscribers: data access subscribers get new objects pushed (S3 notifications / SQS) for SIEM ingestion; query access subscribers query in place with Athena through Lake Formation permissions.
  • Multi-account via a delegated administrator.
"Normalize logs from AWS and our on-premises firewall into one open schema for our SIEM, with the least custom code" → Security Lake (OCSF) with a subscriber. OCSF is new in the C03 guide (skill 3.1.4) and Security Hub now uses it too.

1.2Analyzing, normalizing and correlating logs

ToolBest for
CloudWatch Logs InsightsInteractive queries on log groups (also cross-account). Example: fields @timestamp, userIdentity.arn | filter eventName="ConsoleLogin" and responseElements.ConsoleLogin="Failure" | stats count() by sourceIPAddress
Amazon AthenaSQL over logs in S3 (CloudTrail, VPC Flow Logs, ALB, WAF, Security Lake). Use partitions (or partition projection) to keep scans cheap
Security Hub findingsFilter, group and prioritize findings across accounts
OpenSearch ServiceFull-text search, dashboards and SIEM-style correlation; Security Analytics detection rules
LambdaParse and normalize custom formats on the way in (Firehose transformation)
Amazon Managed GrafanaDashboards correlating CloudWatch, OpenSearch, Athena and third-party data sources
CloudTrail Lake MaintenanceSQL over immutable event data stores, 7 years by default (longer with extended retention); no new customers since Apr 30, 2026. AWS recommends migrating to CloudWatch
Amazon DetectiveVisual investigation; correlation is done for you

1.3Troubleshooting missing logs and silent alerts

Almost every "logs are missing" question is a permissions problem: either the service can't write, or the key policy won't let it encrypt.

SymptomLikely causeFix
Lambda function produces no CloudWatch logsExecution role lacks logs:CreateLogGroup, logs:CreateLogStream, logs:PutLogEventsAttach AWSLambdaBasicExecutionRole (or equivalent)
API Gateway execution logs never appearNo CloudWatch Logs role ARN set at the API Gateway account level, or logging not enabled on the stageSet the role in API Gateway settings; enable logging per stage
CloudWatch agent installed but nothing arrivesInstance profile missing permissions, agent not started, wrong config, or no route to the Logs endpointAttach CloudWatchAgentServerPolicy; check agent log on the instance; add a logs VPC endpoint or NAT
CloudTrail stopped delivering to S3Bucket policy changed, or the KMS key policy doesn't allow CloudTrail, or the key was disabledRestore bucket policy and key-policy statements; check trail status (LatestDeliveryError)
VPC Flow Logs not publishedIAM role trust policy doesn't trust vpc-flow-logs.amazonaws.com, or S3 bucket policy blocks deliveryFix trust policy / bucket policy
CloudFront standard logs missing in S3Legacy standard logging needs ACLs enabled on the target bucket; the bucket has Object Ownership "bucket owner enforced"Enable ACLs on the log bucket, or use the newer logging to CloudWatch Logs / Firehose
WAF logging won't saveDestination name doesn't start with aws-waf-logs-Rename the log group / bucket / stream
ELB access logs failBucket policy doesn't allow the ELB log delivery principal, or bucket uses SSE-KMS (unsupported for ALB logs; use SSE-S3)Correct bucket policy; use SSE-S3
Security Hub CSPM controls show "no data"AWS Config isn't recording that resource type in that RegionEnable Config recording (all resource types)
EventBridge rule for an IAM / console event never firesGlobal-service CloudTrail events (IAM, STS sign-in) are delivered in us-east-1; rule was created in another RegionCreate the rule in us-east-1 (or forward events cross-Region)
Alarm fires but the encrypted SNS topic delivers nothingThe topic's KMS key policy doesn't let CloudWatch / EventBridge use the key (AWS managed aws/sns keys can't be edited)Use a customer managed key with a key policy allowing the service principal
Route 53 health check reports endpoint unhealthy, but it's upSecurity group / NACL blocks Route 53 health checker IP rangesAllow the health checker ranges
Try it

Why is it missing?

Think of your answer, then tap a card to check it.

Sources
Domain 2

Incident response

14% of scored content · Tasks 2.1–2.2: design and test a plan, respond to security events

2.1The incident response lifecycle and runbooks

AWS follows the NIST SP 800-61 lifecycle. Exam questions place you in one phase and ask for the next correct action.

1. Preparation→2. Detection & analysis→3. Containment, eradication & recovery→4. Post-incident activity (lessons learned)↺
Building blockUse it for
Systems Manager Automation runbooksCodified response steps (isolate instance, snapshot volumes, disable access key, revoke public access). AWS provides managed runbooks; you write your own in YAML/JSON. Can run cross-account and cross-Region.
Systems Manager OpsCenterCentral place for OpsItems: collects related findings, alarms and events into one work item, shows related resources and lets responders run the linked runbook from the item.
SageMaker AI notebooks (Jupyter)Interactive investigation playbooks: documented queries (Athena, CloudTrail) and analysis steps that analysts run and record as they go.
Step FunctionsOrchestrate multi-step, branching response (triage → isolate → capture → notify → wait for approval)
EventBridgeTrigger the right runbook from a finding automatically
A playbook says what to do for a type of incident (for humans). A runbook is the how, step by step, ideally as code.

2.1Prepare before anything happens

PreparationHow on AWS
Provision access in advanceIncident-response IAM roles pre-deployed in every account (CloudFormation StackSets), an IR permission set in Identity Center, documented break-glass access. You don't want to be writing IAM policies mid-incident.
Deploy security tools aheadGuardDuty, Security Hub, Detective, Inspector, Macie enabled org-wide; a dedicated forensics account with an isolated VPC, a forensic AMI (built with EC2 Image Builder) and an evidence bucket with Object Lock
Minimize the blast radiusMulti-account separation, SCPs, least privilege, network segmentation, short-lived credentials, encryption with separate keys per data domain
Shield Advanced protectionsProtect internet-facing resources before an attack; associate Route 53 health checks (health-based detection), enable proactive engagement so the Shield Response Team (SRT) calls you, grant the SRT access with an IAM role, enable the WAF Anti-DDoS managed rule group
Turn logging on nowYou can't investigate with logs you never collected: org trail, data events for sensitive buckets, VPC Flow Logs, Resolver query logs
ContactsSet the account's alternate security contact so AWS can reach the right team (for example, about exposed credentials)

2.1Test the plan

MethodWhat it proves
Tabletop exercisePeople know their roles and the playbook makes sense (discussion only)
Game day / simulationThe team and tooling actually work under realistic conditions
AWS Fault Injection Service (FIS)Inject real faults (stop instances, throttle APIs, disrupt network, fail an AZ) with stop conditions tied to CloudWatch alarms; verify that detection and automated response kick in
AWS Resilience HubAssess an application against RTO/RPO targets, recommend improvements, and generate FIS experiments and alarms to validate recovery
GuardDuty sample findingsCreateSampleFindings produces fake findings so you can test the EventBridge → runbook → notification pipeline end to end
"Validate that the automated isolation workflow actually runs, without a real attack" → generate GuardDuty sample findings. "Prove the app meets its RTO when an AZ fails" → FIS experiment (Resilience Hub to set and assess targets).

2.1Automated remediation patterns

Finding / event→EventBridge rule→Lambda, SSM Automation or Step Functions→Fix + notify (SNS) + record (OpsCenter / ticket)
TriggerAutomated response
GuardDuty: EC2 instance talking to a command-and-control serverStep Functions: tag, snapshot, swap to an isolation security group, notify; or the Automated Forensics Orchestrator for Amazon EC2
AWS Health event AWS_RISK_CREDENTIALS_EXPOSED (key found on a public repo)Lambda: deactivate the access key, review CloudTrail for its use, notify the owner
Config rule: S3 bucket made publicConfig automatic remediation with SSM Automation: re-enable Block Public Access
Config rule: security group opens 0.0.0.0/0 on port 22SSM Automation removes the rule
Security Hub custom action "Isolate"Analyst-approved trigger of the same runbook
Regional or zonal impairment / need to move traffic awayAmazon Application Recovery Controller: zonal shift away from an AZ, or routing controls to fail over a Region

Automated Forensics Orchestrator for Amazon EC2

An AWS Solution (Guidance) built on Step Functions. A finding (GuardDuty, Config, third party) reaches Security Hub → EventBridge, which starts the workflow: triage, then in parallel memory acquisition (stored in S3) and disk acquisition (EBS snapshot shared to the forensics account), isolation with a restrictive security group, then automated investigation on a forensic instance. Results are tracked in DynamoDB.

Auto-remediate the boring and certain (public bucket, open SSH). Keep a human in the loop for the destructive or ambiguous (terminating production, deleting users).

2.2Responding to a compromised EC2 instance

The order matters: preserve volatile evidence first, then contain, then investigate a copy, then rebuild.

  1. Capture metadata and tag the instance (for example Quarantine=true); enable termination protection.
  2. Remove it from service: detach from the Auto Scaling group (or set it to Standby) so it isn't replaced and terminated, and deregister from the load balancer.
  3. Isolate: replace its security groups with an isolation security group (no inbound, no outbound or only to forensic tools). Existing tracked connections may survive a security group change, so add a network ACL deny for an immediate cut.
  4. Capture memory while it's running (memory is lost on stop).
  5. Snapshot the EBS volumes and share the snapshots (with a customer managed key) to the forensics account.
  6. Investigate a copy in the isolated forensic environment, never the original.
  7. Eradicate and recover: terminate, relaunch from a known-good, patched AMI, rotate any credentials it held (including its instance role's sessions).
Try it · ordering question

Put the containment steps in order

Click the steps in the order you'd do them. One option doesn't belong at all; leave it out. Then check.

Don't stop or terminate first. Stopping erases memory evidence; terminating can delete the root volume (DeleteOnTermination) and everything with it.

2.2Containing other resource types

CompromisedContainRemember
IAM user access keyDeactivate the key (keep it for evidence), then rotate; review CloudTrail for everything the key didCheck for new users, keys, roles or policies the attacker created (persistence)
IAM role (temporary credentials stolen)Revoke active sessions on the role: IAM adds an inline deny policy with an aws:TokenIssueTime condition, so every session issued before now failsSTS credentials can't be "deleted"; the revoke-sessions deny is how you kill them. Then fix the cause (for example, enforce IMDSv2)
Root userReset password, replace MFA, delete root access keys, check billing and IAM changesWith centralized root access, member accounts may have no root credentials at all
S3 bucket exposed / data exfiltrationTurn on Block Public Access, add a deny bucket policy, rotate exposed credentialsUse Macie to scope what sensitive data was exposed; CloudTrail data events to see what was read
Lambda functionSet reserved concurrency to 0 (stops all invocations), remove triggers, restrict the execution roleCheck the function's code and layers versions for tampering
EKS pod / nodeLabel and isolate the pod with a Kubernetes network policy, cordon the node, revoke the pod's IAM role, capture the nodeGuardDuty EKS Protection and Runtime Monitoring findings tell you which pod
RDS databaseChange master password (or rotate in Secrets Manager), restrict security group, snapshot for evidenceEnable deletion protection
Web application under attackWAF block rules (IP set, rate-based, geo), Shield Advanced SRTCount mode first if false positives would hurt

2.2Forensic artifacts, searching, and validating findings 2.2.3 new

Capture and store evidence

  • Copy relevant logs (CloudTrail, VPC Flow Logs, application and OS logs) plus snapshots and memory images to an evidence bucket in the forensics account: versioning, Object Lock, SSE-KMS with a dedicated key, tight bucket policy.
  • Record SHA-256 hashes of artifacts and keep a chain of custody (who touched what, when).
  • Use CloudTrail log file integrity validation to prove log authenticity.

Search and correlate

QuestionWhere to look
Everything a principal did in the last 30 days across accountsAthena on the org trail (or Security Lake); CloudWatch Logs Insights cross-account; Detective
Which IPs an instance talked toVPC Flow Logs (Athena), Detective
Which domains it resolvedRoute 53 Resolver query logs
Correlate app logs with API activityOpenSearch, Security Lake + Athena, or a SIEM subscriber

Validate findings to assess scope and impact

  • True or false positive? Check finding details (actor, IPs, user agent, resource), compare with known-good behavior; archive known-benign patterns with suppression rules.
  • Scope: which accounts, Regions, principals and resources are involved? Detective finding groups and GuardDuty attack sequences connect related findings.
  • Impact: was data touched (Macie, S3 data events)? Is the vulnerability actually reachable (Inspector network reachability, Security Hub exposure)? What can the compromised identity reach (IAM Access Analyzer, policy review)?

2.2Root cause analysis with Amazon Detective

Detective automatically builds a behavior graph from CloudTrail, VPC Flow Logs, GuardDuty findings and EKS audit logs (up to a year of history), so you can pivot from a finding to the entities involved without writing queries.

  • Finding groups cluster related findings and entities into one investigation.
  • Investigations for IAM users and roles highlight indicators of compromise (unusual geolocation, new API calls, impossible travel).
  • Integrates with GuardDuty and Security Hub ("Investigate in Detective").
  • Root cause method: establish timeline → identify initial access → trace lateral movement and persistence → identify control gaps → fix and update the playbook.
GuardDuty rings the bell. Detective shows you the security camera footage from before the bell rang.
"Least effort to determine the root cause and full scope of a GuardDuty finding" → Amazon Detective. "Run custom SQL across a year of CloudTrail from every account" → Athena on the org trail or Security Lake.

2.2Eradicate and recover

  • Rebuild, don't clean: redeploy from IaC and known-good AMIs/images rather than trusting a cleaned host.
  • Restore data from backups made before the compromise (AWS Backup point-in-time, a logically air-gapped vault if the account itself was compromised).
  • Rotate every secret and key the attacker could have seen (Secrets Manager rotation, new access keys, KMS key rotation if key material is suspected).
  • Patch the entry point, tighten policies, add detections for the technique used.
  • Post-incident review: update playbooks and automation, track action items.
Try it

What's the next step?

Think of your answer, then tap a card to check it.

Sources
Domain 3

Infrastructure security

18% of scored content · Tasks 3.1–3.3: edge, compute workloads, network controls

3.1Edge security strategy: match the control to the threat

ThreatControl
Network/transport DDoS (SYN floods, UDP reflection)Shield Standard (free, automatic, all customers); Shield Advanced for larger attacks, SRT help, cost protection. Put CloudFront / Global Accelerator in front to absorb at the edge
HTTP request floods (layer 7 DDoS)CloudFront caching + AWS WAF rate-based rules + the WAF Anti-DDoS managed rule group
OWASP Top 10 (SQL injection, XSS, bad inputs)AWS WAF AWS Managed Rules: Core rule set, SQL database, Known bad inputs, plus language/OS-specific groups
Scrapers, scanners, botsWAF Bot Control (Common or Targeted), CAPTCHA and Challenge actions
Credential stuffing on a login pageWAF Fraud Control account takeover prevention (ATP)
Fake account sign-upsWAF Fraud Control account creation fraud prevention (ACFP)
Traffic from sanctioned countriesCloudFront geographic restrictions (whole distribution) or WAF geo match rule (finer: per path, with labels, count first)
Users bypassing CloudFront to hit S3 directlyOrigin access control (OAC) + bucket policy allowing only that distribution (aws:SourceArn)
Users bypassing CloudFront to hit the ALBCloudFront adds a secret custom header and the ALB (or WAF on the ALB) requires it; restrict ALB security group to the CloudFront managed prefix list; or use CloudFront VPC origins to keep the ALB private
Premium content for paying users onlyCloudFront signed URLs (one file) or signed cookies (many files), with trusted key groups
Sensitive form fields readable by middle tiersCloudFront field-level encryption (public key encrypts chosen fields at the edge)
IoT devices connecting at scaleAWS IoT Core: X.509 device certificates (mutual TLS) + IoT policies attached to certificates; policy variables such as ${iot:Connection.Thing.ThingName} limit each device to its own topics
Browsers on another site calling your S3 bucketS3 CORS configuration with specific AllowedOrigins and methods. CORS relaxes the browser's same-origin rule; it isn't access control, so keep it narrow and still use bucket policies
Try it

Pick the edge control

3.1AWS WAF in depth

  • Where: a web ACL attaches to CloudFront (global; create in us-east-1), ALB, API Gateway REST APIs, AppSync, Cognito user pools, App Runner, Amplify and Verified Access. Not NLBs or EC2 directly.
  • Rule actions: Allow, Block (custom response possible), Count (test without impact), CAPTCHA (human puzzle), Challenge (silent browser check). Rules run in priority order; the web ACL has a default action.
  • Statements: IP sets, geo match, regex pattern sets, string/size constraints, SQLi and XSS detection, label match, and scope-down statements to narrow managed rules.
  • Labels: a rule can add a label (for example from Bot Control); later rules act on that label. Lets you combine signals.
  • Rate-based rules: count requests per aggregation key over a window; aggregate by IP, forwarded IP header, a header/cookie/query value, label namespace, HTTP method, or JA3/JA4 TLS fingerprint (client fingerprinting, skill 3.1.3). Combine keys for "per session per path" limits.
  • Capacity: each rule costs web ACL capacity units (WCUs); a web ACL has a default allowance you can raise for extra cost.
  • Third-party rules 3.1.4 new: AWS Marketplace managed rule groups from security vendors drop into a web ACL like AWS Managed Rules.
  • Logging to CloudWatch Logs, S3 or Firehose (names start with aws-waf-logs-); redact sensitive fields. WAF logs are a native Security Lake source (normalized to OCSF).
  • Behind a CDN or proxy: inspect the client IP from X-Forwarded-For using the forwarded IP configuration, otherwise you block the proxy.
New managed rules always start in Count mode, check the logs for false positives, then switch to Block. "Test a new rule without affecting users" → Count.

3.1AWS Shield Standard vs Shield Advanced

Shield StandardShield Advanced
CostFree, automaticMonthly subscription per organization (1-year commitment) plus data transfer
ProtectsAll AWS customers against common L3/L4 attacksCloudFront, Route 53 hosted zones, Global Accelerator, ALB, Classic LB, Elastic IPs (EC2, NLB)
DetectionAutomaticTailored per resource; health-based detection using Route 53 health checks; L7 alerts via CloudWatch
HumansNoShield Response Team (SRT) 24/7 (with Business or Enterprise Support); proactive engagement
Money—Cost protection: credits for scaling charges caused by a DDoS attack; AWS WAF included at no extra cost for protected resources
Visibility—Attack diagnostics and history, protection groups
At scale—Deploy protections across accounts with Firewall Manager

For Elastic IPs, Shield Advanced pushes your network ACLs to the AWS border during an attack. For layer 7 it alerts by default rather than auto-blocking, to avoid blocking real users; the WAF Anti-DDoS managed rule group (default since March 26, 2026) mitigates HTTP floods in seconds.

Standard = the free lock on every door. Advanced = a security company on retainer: guards on call, insurance for damage, and a report after every break-in.

3.2Hardened AMIs and container images

  • EC2 Image Builder pipelines build golden AMIs and container images on a schedule: base image → components (install agents, apply CIS or STIG hardening components) → test (including an Inspector vulnerability scan) → distribute to accounts and Regions (and share with the organization). Rebuild when new patches land.
  • Systems Manager keeps configuration in line after launch (State Manager, Patch Manager, Run Command).
  • EC2 Allowed AMIs (account setting, enforceable with a declarative policy) limits which AMI owners/sources instances can launch from.
  • IMDSv2 required (session tokens) blocks SSRF-style credential theft; set the hop limit to 1 if containers on the host shouldn't reach the metadata service.
  • EBS encryption by default per Region.
  • Container images: minimal/distroless bases, non-root users, ECR enhanced scanning (Inspector, continuous), immutable tags, image signing (AWS Signer with Notation) verified at deploy.
Golden AMI = bake security in once, then replace instances. Patching running servers by hand = repainting a moving car.

3.2Roles for compute: who gets which credentials

WorkloadMechanismNotes
EC2Instance profile (container for one IAM role)Credentials served by IMDS, auto-rotated
ECS taskTask role = what your app code can call. Task execution role = what the ECS agent needs (pull from ECR, write logs, fetch secrets for injection)Mixing these up is a favorite question
LambdaExecution role = what the function can call. Resource-based policy = who may invoke the functionS3/SNS/EventBridge triggers need the resource policy
EKS podEKS Pod Identity (associate a role with a service account) or IRSA (OIDC)Never rely on the node role for pods
AWS service acting for youService role (you create, the service assumes) or service-linked role (predefined by the service, you can't edit its permissions)Passing a role to a service needs iam:PassRole
On-premises / other cloudIAM Roles Anywhere (X.509 certificate → temporary credentials)See Domain 4
"The container can pull its image but the app gets AccessDenied writing to DynamoDB" → the task role is missing permissions (the execution role is fine). Restrict who can hand roles to services with iam:PassRole and the iam:PassedToService condition.

3.2Vulnerability scanning and patching

Amazon Inspector

TargetHow
EC2Agent-based via SSM Agent, or agentless (scans EBS snapshots); OS and language packages; network reachability findings (ports reachable from the internet); CIS benchmark assessments
ECR imagesEnhanced scanning on push and continuously as new CVEs appear
LambdaStandard scanning (package dependencies) and code scanning (code-level flaws such as injection, hardcoded secrets)
Code repositoriesCode security: SAST, software composition analysis and IaC scanning for GitHub and GitLab, with findings in pull requests

Inspector scores findings (Inspector score adjusts CVSS for your environment), exports SBOMs, supports suppression rules and runs org-wide via a delegated admin.

GuardDuty for runtime

Inspector finds weaknesses before exploitation; GuardDuty Runtime Monitoring watches EKS, ECS and EC2 for active exploitation (reverse shells, crypto miners, privilege escalation) and Malware Protection scans volumes and S3 uploads.

Patching with Systems Manager Patch Manager

  • Patch baselines define approved patches (auto-approve after N days, by severity/classification).
  • Patch policies (Quick Setup) apply scanning and installing across the whole organization and Regions; or maintenance windows running AWS-RunPatchBaseline.
  • Scan-only vs install; compliance shows in Systems Manager and Security Hub.
  • Continuous validation: Inspector rescans after patching and closes findings; anything still open is a gap.
  • For immutable infrastructure, patch the image (Image Builder) and replace instances instead.

3.2Secure administrative access

OptionInbound port?Public IP / bastion?Pick it when…
Systems Manager Session ManagerNoneNoDefault answer: shell or port forwarding controlled by IAM, sessions logged to S3 / CloudWatch Logs and encrypted with KMS, no SSH keys. Private subnets need VPC endpoints for ssm, ssmmessages, ec2messages
EC2 Instance Connect22 (SSH)Public IP or reachable networkTeams that need real SSH; pushes a one-time public key valid for 60 seconds, controlled by IAM
EC2 Instance Connect Endpoint22 from the endpoint onlyNoSSH/RDP to instances in private subnets with no bastion and no public IP
EC2 serial console——Break-glass troubleshooting of an unbootable instance (can be disabled org-wide with a declarative policy)
Bastion host22YesRarely the best answer now: more to patch and expose
"Administrators need shell access to private instances with no inbound ports, no SSH keys, and a full audit log of commands" → Session Manager with session logging.

3.2Security in the pipeline

Developer IDE: Amazon Q Developer security scans→Pull request: Inspector code security (SAST, SCA, IaC)→Build: cfn-lint, CloudFormation Guard, image build→ECR scan + sign→Deploy with least-privilege cross-account role→Runtime: Inspector + GuardDuty
  • Amazon Q Developer: code reviews in the IDE flag vulnerabilities, secrets and IaC misconfigurations and suggest fixes.
  • Amazon CodeGuru Security Retired: named in the guide (static analysis for security flaws); support ended November 20, 2025. Its job moved to Inspector code security and Q Developer.
  • Lambda code signing: a code signing configuration rejects deployments not signed by an approved AWS Signer profile.
  • Fail the build on critical findings (quality gate); keep pipeline roles minimal; store secrets in Secrets Manager, not build variables.

3.2Guardrails for generative AI apps 3.2.7 new

The guide cites the OWASP Top 10 for LLM Applications (2025 edition). Know each risk and the AWS control that addresses it.

#RiskWhat it isAWS controls
LLM01Prompt injectionInput (direct, or hidden in documents/web pages) overrides the developer's instructionsBedrock Guardrails prompt attack filter; input validation (Lambda); keep untrusted content separate; least-privilege tools
LLM02Sensitive information disclosureModel reveals PII, secrets or confidential dataGuardrails sensitive information filters (block or mask PII, custom regex); Macie on source data; per-user access filters on retrieval
LLM03Supply chainCompromised models, datasets, plugins or librariesUse vetted model providers (Bedrock), SageMaker model registry, Inspector/SBOM for dependencies
LLM04Data and model poisoningTampered training, fine-tuning or RAG dataLock down S3 sources (bucket policies, versioning, Object Lock), validate data, track lineage
LLM05Improper output handlingModel output passed unchecked to SQL, shells or browsersTreat output as untrusted: validate, encode, parameterize
LLM06Excessive agencyAgents with too many tools or permissions act harmfullyLeast-privilege IAM for agent roles; limit tools; human approval for high-impact actions (Bedrock Agents user confirmation / return of control, Step Functions callback)
LLM07System prompt leakageAttackers extract the system prompt (and any secrets in it)Never put secrets in prompts (use Secrets Manager); Guardrails prompt-leakage detection
LLM08Vector and embedding weaknessesUnauthorized access or cross-tenant leaks via the vector storeAccess control and encryption on the vector store; metadata filters per user/tenant in Knowledge Bases
LLM09MisinformationConfident but false answers (hallucinations)Guardrails contextual grounding and Automated Reasoning checks; citations; human review
LLM10Unbounded consumptionAbuse drives runaway cost or denial of serviceAPI Gateway throttling and usage plans, WAF rate limits, max token limits, Bedrock quotas, budgets

Baseline controls for Bedrock, SageMaker AI and Amazon Q

  • Bedrock: your prompts and outputs aren't used to train the base models; reach it privately through PrivateLink interface endpoints; KMS for custom models, agents and knowledge bases; CloudTrail logs API calls and model invocation logging (off by default) records prompts and responses; enforce an approved guardrail with the IAM condition key bedrock:GuardrailIdentifier.
  • SageMaker AI: VPC-only mode, network isolation for training and models (no outbound calls), KMS for volumes and artifacts, inter-container traffic encryption for distributed training, least-privilege execution roles.
  • Amazon Q Business: identity through IAM Identity Center; answers respect the source system's document permissions (ACLs).
Try it · matching question

Match each LLM risk to its best control

3.3Security groups, network ACLs and firewalls

Security groupNetwork ACL
Applies toAn ENI (instance, endpoint, load balancer…)A subnet
StateStateful: return traffic allowed automaticallyStateless: return traffic needs its own rule (ephemeral ports 1024–65535)
RulesAllow onlyAllow and deny
EvaluationAll rules togetherLowest rule number first; first match wins
Best forWorkload micro-segmentation; reference other security groups instead of IPsCoarse subnet guardrails; blocking a specific IP; immediate cut-off in an incident
Try it

Will the HTTPS request get through?

A client on the internet connects to a web server on TCP 443 in a public subnet. Set each control.

Security group inbound
Security group outbound
NACL inbound
NACL outbound

Beyond SGs and NACLs

ControlWhat it adds
AWS Network FirewallManaged stateful + stateless firewall for a VPC: Suricata-compatible IPS rules, domain allow/deny lists (egress filtering by FQDN/SNI), TLS inspection using certificates from ACM, alert/flow/TLS logs. Deploy in a dedicated firewall subnet with route table changes; centrally in an inspection VPC behind Transit Gateway (with appliance mode)
Route 53 Resolver DNS FirewallBlock or alert on DNS queries to bad domains (AWS managed domain lists for malware and botnets, plus your own); stops DNS-based exfiltration and C2 lookups
Gateway Load BalancerInsert third-party firewall appliances transparently (GENEVE)
VPC Block Public AccessAccount/Region setting that blocks internet gateway traffic to and from VPCs (with exclusions); enforce org-wide with a declarative policy
Firewall ManagerDeploy all of the above consistently across accounts (Domain 6)
Security group = bouncer at each door who remembers who came in. NACL = guard at the street gate with a list, who checks people both ways and forgets faces. Network Firewall = airport security that opens the bags.

3.3Hybrid and multi-cloud connectivity

OptionEncrypted?Pick it when…
Site-to-Site VPNYes (IPsec, two tunnels)Quick, encrypted link over the internet; backup for Direct Connect; can authenticate with certificates from Private CA instead of pre-shared keys
Direct ConnectNot by defaultPrivate, consistent bandwidth. Add encryption with:
  ↳ MACsecYes, layer 2Line-rate encryption between your router and the AWS DX router, on supported dedicated connections (10 Gbps and up)
  ↳ IPsec VPN over Direct ConnectYes, layer 3End-to-end IPsec on top of DX (for example on a transit VIF or public VIF), works on any connection size
Client VPNYes (TLS)Remote users need network-level access; authenticate with mutual certificates, Active Directory or SAML; authorization rules per group; consider split tunnel
Verified AccessYesZero-trust, VPN-less access to specific apps: each request checked against identity (Identity Center, OIDC) and device posture (Jamf, CrowdStrike…) with Cedar policies. HTTP(S) apps, and now TCP resources (SSH, RDP, RDS databases) through the Connectivity Client
"Encrypt traffic on an existing 10 Gbps dedicated Direct Connect at line rate" → MACsec. "Encrypt over a 500 Mbps hosted connection" → IPsec VPN over DX. "Contractors need access to one internal web app; replace the VPN; check device posture" → Verified Access.

3.3Network segmentation

DirectionMeaningControls
North/southIn and out of your network (internet, on-premises)Public ALB + WAF for ingress; private subnets; NAT gateway for egress; centralized egress VPC with Network Firewall domain filtering; VPC Block Public Access
East/westBetween workloads, subnets and VPCsSecurity groups referencing each other; Transit Gateway route tables to separate prod, dev and shared; Network Firewall in an inspection VPC; VPC Lattice auth policies for service-to-service
  • Isolated subnets: no route to an internet gateway or NAT. Reach AWS services through VPC endpoints: gateway endpoints (S3, DynamoDB, free) or interface endpoints (PrivateLink, most services).
  • Endpoint policies limit what can be done through an endpoint (for example only your org's buckets), and bucket policies can require aws:SourceVpce so data is only reachable from your network.
  • Three-tier pattern: public subnet (ALB) → private app subnet → isolated data subnet, each with its own security groups.

3.3Finding unnecessary network access

ToolFinds
Network Access AnalyzerPaths that violate your intent, defined as network access scopes: "no path from an internet gateway to the database subnets", "only the firewall subnet reaches the IGW"
Inspector network reachabilityEC2 ports open and reachable from the internet (and which route makes them reachable)
Verified AccessReplaces broad VPN network access with per-app access, removing unneeded paths
Firewall Manager security group audit policiesOverly permissive or unused security groups across accounts, with remediation
AWS Config / Security Hub CSPMRules like restricted-ssh, unrestricted ports on security groups
IAM Access AnalyzerResources reachable through policies (not network) from outside
"Continuously verify that no production database is reachable from the internet, across many VPCs" → Network Access Analyzer. Reachability Analyzer (point-to-point connection troubleshooting) was dropped from the C03 blueprint.
Sources
Domain 4 · largest

Identity and access management

20% of scored content · Tasks 4.1–4.2: authentication and authorization

4.1Identity solutions for humans, apps and systems

WhoAnswerKey points
Workforce users across many accountsAWS IAM Identity CenterOne sign-in portal for all accounts and apps. Identity source: its own directory, Active Directory (Managed Microsoft AD or AD Connector), or an external IdP (Okta, Entra ID) via SAML 2.0 with automatic provisioning via SCIM. Permission sets become IAM roles in each assigned account. Supports MFA and ABAC from IdP attributes
Customers of your web / mobile appAmazon Cognito user poolsSign-up/sign-in, social and SAML/OIDC federation, MFA, threat protection (adaptive authentication, compromised-credential checks), managed login pages; issues ID, access and refresh JWTs
App users who need AWS credentials directlyCognito identity poolsExchange a token (user pool, social, SAML, OIDC) for temporary IAM credentials; authenticated vs unauthenticated (guest) roles; role mapping by claims
Corporate directory in AWSAWS Directory ServiceManaged Microsoft AD (real AD, trusts with on-premises AD), AD Connector (proxy to on-premises AD, nothing cached in AWS), Simple AD (basic Samba)
Applications on AWSIAM roles (instance profile, task role, execution role, Pod Identity)Never long-term access keys in code
Servers and workloads outside AWSIAM Roles AnywhereX.509 certificates → temporary credentials (see 4.2)
CI/CD systems (GitHub Actions, GitLab)IAM OIDC identity provider + role with a trust policy scoped to the repo/branchNo stored keys; AssumeRoleWithWebIdentity

Multi-factor authentication

  • Types: passkeys and FIDO2 security keys (phishing-resistant), virtual authenticator apps (TOTP), hardware TOTP tokens. A user or root can register several MFA devices.
  • AWS now enforces MFA for root user console sign-in (rolled out starting with management and standalone accounts). For member accounts, the best practice is to centralize root access and remove root credentials entirely (Domain 6).
  • Require MFA for sensitive API calls with "Bool": {"aws:MultiFactorAuthPresent": "true"} or aws:MultiFactorAuthAge. For IAM users calling the CLI, they get an MFA-backed session with sts:GetSessionToken.
User pools = who are you (sign-in, JWTs). Identity pools = which AWS keys do you get. Identity Center = the front door for employees.

4.1Temporary credentials: STS and presigned URLs

APIUsed byNotes
AssumeRoleUsers, roles, services; cross-account accessSession 15 minutes up to the role's maximum (1–12 hours). Role chaining (a role assuming another role) is capped at 1 hour. Supports ExternalId, session tags, session policies, SourceIdentity
AssumeRoleWithSAMLSAML federation into an account's rolesIdP sends Role and Provider ARNs in the SAML assertion
AssumeRoleWithWebIdentityOIDC: CI pipelines, EKS IRSA, mobile appsTrust policy conditions on the token's aud and sub
GetSessionTokenIAM usersGet MFA-authenticated temporary credentials
GetFederationTokenLegacy custom identity brokerRarely the right answer today
  • External ID: a secret the third party passes when assuming your role, preventing the confused deputy problem (someone tricking the vendor into using its access on your account).
  • Session tags carry attributes (department, project) into the session for ABAC; passing them needs sts:TagSession in the trust policy. Source identity (sts:SetSourceIdentity) records the original human through role chains and can't be changed later.
  • S3 presigned URLs grant time-limited access to one object using the signer's permissions. Maximum 7 days with SigV4 and long-term keys, but a URL signed with temporary credentials expires when those credentials expire (often much sooner). If the signer loses permission, the URL stops working.
  • CloudFront signed URLs/cookies are the edge equivalent for content behind CloudFront.
"A presigned URL stops working after an hour even though it was set for 12 hours" → it was signed with temporary role credentials that expired. Sign with credentials that live long enough (or generate on demand).

4.1Troubleshooting authentication

SymptomLikely cause
AccessDenied calling sts:AssumeRoleTarget role's trust policy doesn't name the caller (or the caller's account), a condition fails (sts:ExternalId, MFA, aws:PrincipalOrgID), or the caller's own identity policy / an SCP doesn't allow sts:AssumeRole
AssumeRole works without tags but fails with session tagsTrust policy lacks sts:TagSession
Identity Center user signs in but sees no account / roleNo assignment for the user or group, or the permission set wasn't re-provisioned after a change
Identity Center users from Okta missing or staleSCIM provisioning token expired or mapping error
ABAC through Identity Center not workingAttributes for access control not enabled/mapped
SAML login failsWrong ACS URL / audience, certificate mismatch, clock skew, missing Role attribute
Cognito sign-in errorsUser not confirmed, wrong app client / callback URL, missing client secret, expired refresh token
AD Connector can't authenticateService account permissions, DNS resolution or security group ports to the domain controllers

Start with CloudTrail: ConsoleLogin events (success/failure, MFA used), AssumeRole* events and their errorCode/errorMessage. Identity Center sign-in events are logged too.

4.2Policy types and evaluation logic

Policy typeAttached toGrants?
Identity-based (managed / inline)User, group, roleYes
Resource-basedBucket, key, queue, topic, secret, function, role trust policy…Yes (and can grant cross-account)
Permissions boundaryUser or roleNo, sets a maximum
Service control policy (SCP)Org root, OU, accountNo, limits principals in member accounts
Resource control policy (RCP)Org root, OU, accountNo, limits access to resources in member accounts
Session policyPassed at AssumeRole / federationNo, narrows the session
VPC endpoint policyEndpointNo, limits what passes through it
ACLs (S3 legacy)Bucket / objectYes; disable with Object Ownership "bucket owner enforced"
Explicit deny anywhere?→SCPs allow?→RCPs allow?→Resource-based policy?→Identity-based policy?→Permissions boundary?→Session policy?
Try it

Will this request be allowed?

A role in account A calls s3:GetObject on a bucket. Set each policy layer, then read the verdict.

SCP on the caller's account
RCP on the bucket's account
Identity policy on the role
Bucket policy (resource-based)
Permissions boundary on the role
Session policy (passed at AssumeRole)
VPC endpoint policy (request goes through an S3 endpoint)
SCPs, RCPs, boundaries, session and endpoint policies are fences: they never grant, only limit. Identity and resource policies are keys. You need a key and must be inside every fence. Cross-account, you need a key from both sides.

4.2Cross-account access and authorization services

Two ways to cross accounts

PatternHowWhen
Role assumptionTarget account role with a trust policy naming the source; source principal needs sts:AssumeRoleWorks for every service; the caller gives up its own permissions while in the role
Resource-based policyBucket/key/queue/topic/secret/function policy names the other account or principal; the caller's identity policy must also allow itCaller keeps its own identity (useful for copying data between accounts)
  • Scope trust and resource policies with aws:PrincipalOrgID or aws:PrincipalOrgPaths instead of listing accounts.
  • When an AWS service is the principal (for example cloudtrail.amazonaws.com), add aws:SourceArn / aws:SourceAccount to prevent the cross-service confused deputy.
  • KMS cross-account needs the key policy to allow the other account and an IAM policy in that account allowing its principals.
  • IAM paths (for example /security/, /app-team-a/) organize roles and users so policies can target groups: arn:aws:iam::*:role/security/*.

Authorization services

ServiceWhat it authorizes
Amazon Verified PermissionsFine-grained permissions inside your application ("can Alice approve invoice 42?"), written in the Cedar policy language in a central policy store; supports RBAC and ABAC; can use Cognito tokens as identity and back an API Gateway Lambda authorizer
IAM Roles AnywhereWorkloads outside AWS: register your CA (AWS Private CA or your own) as a trust anchor, create a profile (which roles, session policies), roles trust rolesanywhere.amazonaws.com; the credential helper signs with the workload's certificate. Revoke with a CRL. Conditions can match certificate attributes
IAM role trust policiesWho can become the role, under which conditions
Resource policiesWho can use a specific resource, including from other accounts
"Application-level permissions managed outside the code, auditable, written as policies" → Verified Permissions (not IAM; IAM authorizes AWS API calls, not your app's business actions).

4.2ABAC and RBAC

RBAC (role-based)ABAC (attribute-based)
Access decided byWhich role/group you're inMatching tags/attributes on the principal and the resource
Scales byCreating more roles and policiesTagging; the policy stays the same
Good forStable job functions; auditors like itMany projects/teams that change often
{
  "Effect": "Allow",
  "Action": ["ec2:StartInstances", "ec2:StopInstances"],
  "Resource": "*",
  "Condition": {
    "StringEquals": { "aws:ResourceTag/project": "${aws:PrincipalTag/project}" }
  }
}
  • aws:PrincipalTag/key: tag on the caller (from the role, or session tags from the IdP / Identity Center attributes).
  • aws:ResourceTag/key: tag on the existing resource.
  • aws:RequestTag/key and aws:TagKeys: tags in the create/tag request (force new resources to carry the right tags).
  • Protect the tags: deny TagResource/UntagResource on the controlling tag keys, or users can re-tag their way into access.
RBAC = job title badges. ABAC = colored wristbands: you get into any room painted your color, and new rooms just need paint, not new badges.

4.2Least privilege in practice

ToolUse
Permissions boundarySafe delegation: developers may create roles only if they attach the approved boundary. Enforce with a condition on iam:CreateRole / iam:PutRolePermissionsBoundary: "iam:PermissionsBoundary": "arn:…:policy/DevBoundary", and deny removing it
Session policyNarrow a single session (for example a broker giving a user access to one prefix)
IAM Access Analyzer policy generationBuilds a least-privilege policy from the actions a role actually used in CloudTrail
Last accessed informationService- and action-level "last used" data to remove unused permissions
Condition keysaws:SourceIp, aws:SourceVpc, aws:SourceVpce, aws:RequestedRegion, aws:SecureTransport, aws:CalledVia, aws:ViaAWSService, aws:PrincipalOrgID, aws:ResourceOrgID
aws:SourceIp doesn't contain your client's IP when the request comes through a VPC endpoint or from an AWS service calling on your behalf, so an IP-restricting deny can break those. Combine with aws:ViaAWSService or use aws:SourceVpce for private traffic. Also watch NotAction + Allow, which grants everything except the listed actions.

4.2Analyzing authorization failures and unintended access

ToolAnswers
AccessDenied error messagesMost services now say which policy type denied the call ("…with an explicit deny in a service control policy", "…because no identity-based policy allows…")
CloudTrailerrorCode: AccessDenied with the principal, action and resource
sts:DecodeAuthorizationMessageDecodes the encoded failure message some services (such as EC2) return
IAM Policy SimulatorTest identity-based policies (and boundaries, and the SCPs that apply) against actions/resources/conditions without making real calls
IAM Access Analyzer: external accessResources (S3, KMS, IAM roles, Lambda, SQS, Secrets Manager, snapshots…) shared outside your zone of trust (account or organization). Free
IAM Access Analyzer: internal accessWhich principals inside your org can reach chosen critical resources (uses automated reasoning over all policy types)
IAM Access Analyzer: unused accessUnused roles, access keys, passwords and permissions (paid per role/user)
Access Analyzer policy validation and custom policy checksLint policies; in pipelines, CheckNoNewAccess, CheckAccessNotGranted, CheckNoPublicAccess block risky policy changes before deploy
Try it

Why was it denied (or allowed)?

Think of your answer, then tap a card to check it.

Correcting unintended permissions

  1. Find: Access Analyzer findings, last-accessed data, Security Hub IAM controls.
  2. Contain: revoke sessions, deactivate keys, add a deny if urgent.
  3. Fix: remove or narrow the grant (resource policy, trust policy, identity policy); add an SCP/RCP guardrail so it can't recur.
  4. Verify: re-run the analyzer (findings move to resolved), archive intended access with archive rules.
Sources
Domain 5

Data protection

18% of scored content · Tasks 5.1–5.3: data in transit, data at rest, secrets and key material

5.1Requiring encryption in transit

WhereHow to require TLS
Application Load BalancerHTTPS listener with an ACM certificate and a security policy that sets TLS versions and ciphers (for example a TLS 1.3/1.2 policy, FIPS policies, post-quantum hybrid policies); HTTP listener that only redirects to HTTPS; optional mutual TLS with a trust store
Network Load BalancerTLS listener (terminate) or TCP pass-through to targets that terminate TLS
End-to-endRe-encrypt from the load balancer to targets (HTTPS target group); targets can use self-signed or Private CA certificates
CloudFrontViewer protocol policy (redirect HTTP to HTTPS or HTTPS only), minimum TLS security policy for viewers, origin protocol policy HTTPS only
API GatewayHTTPS only; choose a minimum TLS version on custom domains; mutual TLS on custom domains
S3Bucket policy denying requests where aws:SecureTransport is false; s3:TlsVersion to require a minimum version
RDS / AuroraForce SSL with a parameter group (rds.force_ssl for PostgreSQL/SQL Server, require_secure_transport for MySQL/MariaDB) and the RDS CA bundle on clients
ElastiCache, OpenSearch, Redshift, MSKEnable in-transit encryption (and node-to-node encryption for OpenSearch)
{
  "Effect": "Deny",
  "Principal": "*",
  "Action": "s3:*",
  "Resource": ["arn:aws:s3:::example-bucket", "arn:aws:s3:::example-bucket/*"],
  "Condition": { "Bool": { "aws:SecureTransport": "false" } }
}

5.1Secure and private access to resources

OptionWhat it doesPick it when…
Gateway VPC endpointRoute-table entry to S3 or DynamoDB; freePrivate access to S3/DynamoDB from a VPC (not from on-premises)
Interface VPC endpoint (PrivateLink)ENI with private IPs for an AWS service; reachable from on-premises over VPN/DXMost AWS APIs privately (KMS, Secrets Manager, SSM, STS, Bedrock…); also S3 from on-premises
PrivateLink endpoint serviceExpose your service behind an NLB to other VPCs/accountsOne-way private access to an internal or SaaS service; overlapping CIDRs
Endpoint policiesLimit which principals, actions and resources go through the endpointStop data exfiltration to buckets outside your org
Client VPNManaged OpenVPN-based remote accessUsers need network-level access
Verified AccessZero-trust per-app accessUsers need specific apps, no VPN, device checks
"Data in a bucket must only be reachable from our VPC" → bucket policy deny unless aws:SourceVpce equals your endpoint (careful: this also blocks the console and admin access from outside). "Instances may only write to our org's buckets" → endpoint policy with aws:ResourceOrgID.

5.1Inter-resource encryption in transit 5.1.3 new

ServiceHow traffic between nodes/containers is encrypted
Nitro SystemSupported Nitro instance types automatically encrypt traffic between each other (AEAD, 256-bit) with no performance impact, when they're in the same Region and the same VPC or peered VPCs. Not encrypted when the traffic passes through an intermediate hop such as a transit gateway or load balancer
Amazon EMRA security configuration enables in-transit encryption (TLS certificates from a PEM zip in S3 or a custom certificate provider) for node-to-node and application traffic (Spark, Hadoop, Presto), plus at-rest encryption for local disks and EMRFS on S3
Amazon EKSPod-to-pod: Nitro encryption between supported nodes, or a service mesh with mTLS (for example Istio), or VPC Lattice; control plane traffic is TLS. Kubernetes API data at rest gets default envelope encryption with KMS (AWS-owned key by default, or your own key)
SageMaker AIInter-container traffic encryption for distributed training jobs (an option on the job; slower training), network isolation, KMS for volumes and output
OpenSearch ServiceNode-to-node encryption option
Nitro encryption is free and automatic, but only on the direct road (same VPC or peered). Take a detour through a transit gateway or load balancer and you're back to doing TLS yourself.

5.2Choosing how to encrypt data at rest

KMS key types

KeyControlNotes
AWS ownedAWS, sharedDefault for some services; free; you can't see, audit or manage it
AWS managed (aws/s3, aws/ebs…)AWS, in your accountRotated every year automatically; you can't edit the key policy, rotate on demand, or use it cross-account
Customer managedYouYour key policy and grants, cross-account use, enable/disable, scheduled deletion (7–30 days), automatic rotation (90–2,560 days, default 365) and on-demand rotation
  • Key policy is the root of access: unless the key policy allows the account (the default "enable IAM policies" statement), IAM policies alone can't grant use of the key. Grants give temporary, programmatic permissions (used by services like EBS).
  • Envelope encryption: KMS generates a data key; the plaintext data key encrypts the data and is discarded; the encrypted data key is stored with the data.
  • Encryption context: non-secret key-value pairs bound to the ciphertext (additional authenticated data); logged in CloudTrail and usable in key policy conditions (kms:EncryptionContext:…).
  • Useful conditions: kms:ViaService (only when called through, say, S3), kms:CallerAccount, kms:GrantIsForAWSResource.
  • Key specs: symmetric AES-256 (default), asymmetric RSA/ECC (encrypt or sign/verify), HMAC, and ML-DSA post-quantum signing keys.

KMS vs CloudHSM vs custom key stores

OptionWho holds the key materialPick it when…
AWS KMS (standard)KMS HSMs (FIPS 140-3 validated), multi-tenant serviceDefault: integrated with 100+ services, cheapest, least effort
CloudHSM key storeYour CloudHSM clusterKMS integration, but keys must live in single-tenant HSMs you control
External key store (XKS)Your HSM outside AWS, via an XKS proxyRegulation says keys must never be in AWS; you accept latency and availability risk
AWS CloudHSM directlyYour single-tenant HSMs (FIPS 140-3 Level 3) in your VPC; you manage users (crypto officers, crypto users)Apps need standard HSM APIs (PKCS#11, JCE, Microsoft CNG), SSL/TLS offload, Oracle/SQL Server TDE, or sole key custody. Run 2+ HSMs across AZs

Server-side vs client-side

OptionNotes
SSE-S3Default for all new S3 objects; AWS manages keys
SSE-KMSAudit key use in CloudTrail, separate permissions to decrypt, customer managed keys; enable S3 Bucket Keys to cut KMS request costs sharply
DSSE-KMSTwo independent layers of encryption for compliance needs
SSE-CYou supply the key with every request (HTTPS required); AWS never stores it
Client-sideEncrypt before sending (AWS Encryption SDK, S3 Encryption Client, AWS Database Encryption SDK); AWS only ever sees ciphertext; needs code changes
Try it

Pick the encryption key option

Encrypting an existing unencrypted RDS instance or EBS volume: snapshot → copy the snapshot with encryption → restore. You can't flip encryption on in place. Sharing an encrypted snapshot cross-account requires a customer managed key shared via its key policy.

5.2Protecting data integrity

ControlWhat it guarantees
S3 Object LockWORM: object versions can't be deleted or overwritten for a retention period. Requires versioning. Can set a bucket default retention
  ↳ Governance modeUsers with s3:BypassGovernanceRetention can shorten or remove retention. For "protect against accidents, admins can override"
  ↳ Compliance modeNo one, including the root user, can delete or shorten retention until it expires. For regulatory WORM (SEC 17a-4 style)
  ↳ Legal holdNo expiry date; stays until removed (needs s3:PutObjectLegalHold). For litigation
S3 Glacier Vault Lock MaintenanceLocks a vault policy (for the original Glacier vault service) after a 24-hour test window; immutable after completion. New customers use Object Lock instead
Versioning + MFA DeleteRecover overwritten/deleted objects; MFA Delete (root user, CLI only) required to permanently delete versions or change versioning
ChecksumsS3 verifies uploads with checksums (CRC32, CRC32C, CRC64NVME, SHA-1, SHA-256)
CloudTrail log file validationSigned digest files prove logs weren't changed
Digital code signingAWS Signer for Lambda code signing and container images; KMS asymmetric keys for sign/verify of anything; HMAC keys for message authentication
File validation in transferDataSync verifies data integrity after transfer
Try it

Which Object Lock setting?

5.2Lifecycle management and retention

DataToolTypical rule
S3 objectsS3 LifecycleTransition to Standard-IA / Glacier storage classes, expire current or noncurrent versions, abort incomplete multipart uploads. Lifecycle can't delete versions under Object Lock retention
EFS filesEFS lifecycle policiesMove files not accessed for N days to Infrequent Access or Archive; move back to Standard on access
FSx for LustreBackup policiesAutomatic daily backups for persistent file systems with a retention period you set; user-initiated backups kept until deleted; or use AWS Backup
EBS snapshots, AMIsData Lifecycle Manager (DLM)Scheduled snapshots by tag, retention count/age, cross-Region and cross-account copy
Everything AWS Backup supportsBackup plansFrequency, retention, lifecycle to cold storage, copy to other Regions/accounts
LogsCloudWatch Logs retention, S3 lifecycle on log bucketsKeep what compliance needs, delete the rest

5.2Secure backups, replication and ransomware protection

  • AWS Backup: central backup plans across services and accounts; backup policies in Organizations apply them org-wide; cross-account and cross-Region copies; restore testing proves backups work; Backup Audit Manager reports on compliance with your backup controls.
  • Backup Vault Lock: governance mode (privileged users can change it) or compliance mode (after a grace time of at least 72 hours, nobody can change or delete it, and recovery points can't be deleted before retention ends).
  • Logically air-gapped vault: a secondary vault with compliance-mode lock built in, encrypted by default with an AWS owned key (or your key), shareable through RAM (or multi-party approval) so another account can restore directly even if the source account is lost.
  • Data Lifecycle Manager for EBS snapshots/AMIs, including cross-account copy to a vault account.
  • DataSync: encrypts in transit with TLS, verifies integrity, uses IAM roles, and can run through VPC endpoints.
  • S3 Replication to a bucket in a separate, locked-down account (replica owner override), with KMS keys in the destination and Object Lock replicated.
Production account→AWS Backup copy (cross-account, cross-Region)→Vault account: Vault Lock compliance mode / logically air-gapped vault→Restore testing on a schedule
"Backups must survive even if an attacker gets admin (or root) in the production account" → copy to a separate account with Vault Lock in compliance mode or a logically air-gapped vault. Same-account backups and governance mode don't meet this.

5.3Managing and rotating secrets

Secrets ManagerParameter Store SecureString
RotationBuilt in: managed rotation for RDS/Aurora/Redshift/DocumentDB (and RDS-managed master passwords), or a rotation Lambda; schedules as often as every 4 hoursNone built in (write your own)
Cross-RegionReplicate secrets to other RegionsNo
Cross-accountResource-based policy + customer managed KMS keyLimited (advanced parameters can be shared through RAM)
CostPer secret per month + API callsStandard parameters free
Pick it forDatabase credentials, API keys that must rotateConfig values and simple secrets with no rotation need
  • Rotation strategies: single user (change the password in place; brief failures possible) vs alternating users (two users, swap between them; no downtime).
  • Rotation Lambdas in a VPC need a route to the Secrets Manager endpoint and to the database.
  • Reference secrets without exposing them: CloudFormation dynamic references, ECS/EKS secret injection, Lambda extension / caching client.
  • Resource policies: block public sharing (BlockPublicPolicy); restrict to a VPC endpoint with aws:SourceVpce.
"Rotate database passwords automatically every 30 days with the least operational overhead and no app downtime" → Secrets Manager managed rotation (alternating users strategy).

5.3Imported key material and external key stores

AWS-generated vs imported key material 5.3.3 new

AWS-generated (origin AWS_KMS)Imported (origin EXTERNAL)
Who generates itKMS, inside its HSMsYou, in your own key infrastructure; wrapped with a KMS-provided public key and imported
DurabilityAWS keeps it durableYou must keep a copy; if it's deleted or expires in AWS, you must re-import the same material
ExpirationNeverOptional expiration date; KMS deletes it then
Instant removalOnly scheduled deletion of the whole key (7–30 days)DeleteImportedKeyMaterial makes the key unusable immediately (you can re-import later)
RotationAutomatic (90–2,560 days) or on demandOn demand only: import new material (state PENDING_ROTATION), then call RotateKeyOnDemand. Old material stays for decrypting
Key typesAllSymmetric, asymmetric and HMAC keys
Multi-RegionReplicas share material automaticallyImport the same material into the primary and each replica; then rotate on the primary
  • On-demand rotation is limited to 25 times per key and works for symmetric encryption keys (not asymmetric, HMAC or keys in custom key stores).
  • External key store (XKS): KMS keys whose material stays in your HSM outside AWS; every encrypt/decrypt goes through your XKS proxy. "Hold your own key": you can cut AWS off instantly, but your outage becomes their outage.
  • CloudHSM key store: material generated and kept in your CloudHSM cluster; KMS keys in custom key stores don't support automatic rotation.
Imported key material = bringing your own lock cylinder: AWS uses it, but if you pull it out, every door stops opening, and AWS has no spare.

5.3Masking sensitive data 5.3.4 new

FeatureHow it works
CloudWatch Logs data protection policiesAttach to a log group or the whole account. Managed data identifiers (credit cards, emails, AWS keys, names…) and custom identifiers detect sensitive data as it's ingested and mask it in queries, Insights, subscriptions and exports. Only principals with logs:Unmask see the original. Audit findings can go to S3, CloudWatch Logs or Firehose
SNS message data protection MaintenanceData protection policy on a topic to audit, de-identify (mask or redact) or deny messages containing PII/PHI, inbound (publish) or outbound (per subscription). No new customers since April 30, 2026
Bedrock Guardrails sensitive information filtersMask or block PII in prompts and model responses
MacieDiscovers (doesn't mask) sensitive data at rest in S3
"Developers must troubleshoot from application logs, but must not see customer card numbers; the security team must still be able to see them" → CloudWatch Logs data protection policy with logs:Unmask granted only to the security team.

5.3Keys and certificates across Regions 5.3.5 new

KMS multi-Region keys

  • A primary key and replicas in other Regions share the same key ID and key material, so data encrypted in one Region decrypts in another without a cross-Region call.
  • Each replica is its own resource with its own key policy, grants, aliases and tags; rotation is controlled on the primary and synced.
  • Use for DR, active-active apps, DynamoDB global tables with client-side encryption, and signatures verified in many Regions. Don't use them by default: single-Region keys keep data more isolated.

Certificates

ServiceUseRegional notes
ACM public certificatesFree TLS certificates for ALB, NLB, CloudFront, API Gateway, with automatic renewal (DNS validation renews hands-free)Certificates are Regional; request in each Region you use. CloudFront requires the certificate in us-east-1. Exportable public certificates are available (paid) for use outside integrated services
AWS Private CAYour own private CA hierarchy (root → subordinates) for internal TLS, mTLS, IoT devices, Roles Anywhere trust anchors, VPN authenticationA CA lives in one Region; create CAs (for example a subordinate per Region) where you need them; share CAs across accounts with RAM; short-lived certificate mode (validity up to 7 days) is cheaper and avoids revocation; CRL and OCSP for revocation
ACM private certificatesCertificates issued from Private CA but managed (and renewed) by ACMUse on integrated services

Rotation and deletion cheat sheet

KeyRotation
AWS managedAutomatic every year; can't change
Customer managed symmetric (AWS material)Optional automatic, 90–2,560 days; on demand up to 25 times; key ID and ARN unchanged, old material kept to decrypt
Imported materialOn demand after importing new material
Asymmetric, HMAC, custom key store keysNo automatic rotation; create a new key and move the alias
DeletionSchedule with a 7–30 day waiting period; alarm on any attempted use of a key pending deletion; prefer disabling if unsure
Sources
Domain 6

Security foundations and governance

14% of scored content · Tasks 6.1–6.3: accounts, secure deployment, compliance

6.1AWS Organizations and the account structure

  • Accounts are the strongest isolation boundary. Group them into organizational units (OUs) and attach policies at the root, OU or account.
  • Enable all features (not just consolidated billing) to use policies and integrations.
  • The management account pays the bill and owns the org; keep workloads out of it, because SCPs and RCPs don't apply to it.
  • Trusted access lets a service work across the org; delegated administrator moves day-to-day admin of that service out of the management account.

AWS Security Reference Architecture (SRA) accounts

AccountPurpose
ManagementOrg, billing, minimal access
Security ToolingDelegated admin for GuardDuty, Security Hub, Inspector, Macie, Detective, Access Analyzer, Firewall Manager, root access management
Log ArchiveOrg trail, Config history, VPC Flow Logs; locked down, immutable
NetworkTransit Gateway, inspection VPC, centralized egress, DNS
Shared servicesDirectory, CI/CD tooling
Workloads (prod / non-prod)Applications
ForensicsIsolated investigation environment

Common OUs: Security, Infrastructure, Workloads (Prod/Non-prod), Sandbox, Policy Staging, Suspended.

6.1AWS Control Tower

  • Sets up a landing zone: management, Log Archive and Audit accounts, Identity Center, organization-wide CloudTrail and Config.
  • Controls (formerly guardrails) come in three behaviors: preventive (SCPs/RCPs: stop it), detective (Config rules: report it), proactive (CloudFormation Hooks: block noncompliant resources before they're created). And three guidance levels: mandatory, strongly recommended, elective.
  • Optional and custom controls: enable more from the Control Catalog; add your own with Customizations for Control Tower (CfCT: your SCPs and CloudFormation) or Account Factory for Terraform (AFT).
  • Account Factory vends new accounts that are compliant from the first minute.
  • Existing environments: register existing OUs and enroll existing accounts to bring them under governance; drift detection flags manual changes to the landing zone.
  • Built-in Region deny control restricts which Regions can be used.
"New accounts must be created with logging, guardrails and SSO already in place, with the least effort" → Control Tower Account Factory. "Block a noncompliant CloudFormation resource before it's provisioned" → proactive control (CloudFormation Hook).

6.1Organization policies

PolicyControlsTypical use
Service control policy (SCP)Maximum permissions of IAM users and roles (including member-account root) in member accountsDeny leaving the org; deny disabling CloudTrail, GuardDuty, Config; deny unapproved Regions; deny root user actions; require tags
Resource control policy (RCP)Maximum permissions on resources in member accounts, whoever calls (supported services include S3, STS, KMS, SQS and Secrets Manager, with more added over time)Data perimeter: no access from outside the org; require TLS; enforce encryption standards on resources
Declarative policyEnforces a service configuration (desired state) directly in the service, not an API permission. Today: EC2-related settingsRequire IMDSv2 defaults, block public AMI and EBS snapshot sharing, VPC Block Public Access, disable serial console, Allowed AMIs; custom error message; account status report
AI services opt-out policyWhether AWS AI services may store and use your content to improve the serviceOpt every account out in one policy
Tag policyStandard tag keys and allowed valuesConsistent tagging (6.2.2)
Backup policyAWS Backup plans applied org-wideEvery prod account backs up the same way
{
  "Sid": "DenyOutsideApprovedRegions",
  "Effect": "Deny",
  "NotAction": ["iam:*", "organizations:*", "sts:*", "cloudfront:*", "route53:*", "support:*"],
  "Resource": "*",
  "Condition": { "StringNotEquals": { "aws:RequestedRegion": ["eu-west-1", "eu-central-1"] } }
}

SCPs are inherited down the tree, and an action must be allowed at every level (root, each OU, account). The default FullAWSAccess SCP makes "deny list" strategies easy; removing it means building an allow list.

SCP = rules for your people. RCP = rules for your stuff. Declarative policy = factory settings that can't be changed, even by an admin, and even through APIs launched next year.
Try it

Which control fits?

6.1Delegated administrators and central security services

From the management account, register the Security Tooling account as delegated administrator for each service, then turn on auto-enable so every new account is covered.

ServiceCentral capability
GuardDuty, Inspector, Macie, DetectiveDelegated admin, auto-enable members, central findings
Security HubDelegated admin, central configuration policies, cross-Region aggregation
AWS ConfigDelegated admin for org conformance packs, Config rules and aggregators
IAM Access AnalyzerOrganization-wide analyzers
Firewall ManagerFirewall Manager administrator account
Security LakeDelegated admin collects from all accounts
CloudTrailDelegated admin can manage the org trail
IAM (root access management)Delegated admin for centralized root access
AWS Backup, IAM Identity CenterDelegated administration
"Least effort to make sure GuardDuty is on in every current and future account" → delegated admin + auto-enable for all accounts (in each Region). Not a Lambda that enables it account by account.

6.1Root user credentials and break-glass access

Centralized root access for member accounts

  • Root credentials management: from the management or delegated admin account, remove member accounts' root passwords, access keys, signing certificates and MFA, prevent root password recovery, and create new accounts without root credentials.
  • Root sessions: when a task truly needs root (for example unlocking an S3 bucket or SQS queue policy that denies everyone), take a short-lived, task-scoped root session from the central account instead.
  • AWS SRA: centralize, remove member root credentials, delegate to the Security Tooling account.

Root user for the management account (and any account keeping root)

  • Strong password, MFA (FIDO2 / passkey preferred; register more than one), no access keys.
  • Alert on every root sign-in (EventBridge rule on ConsoleLogin with userIdentity.type = Root).
  • In member accounts, an SCP can deny actions by the root user.

Break-glass procedure

  • Emergency access that still works if the IdP or Identity Center fails: a small number of tightly controlled IAM users or roles with hardware MFA, credentials sealed and split among custodians.
  • Alarm on any use, review afterwards, rotate after use, test regularly.

6.2Infrastructure as code, securely and consistently

ToolRole
CloudFormation StackSetsDeploy one template to many accounts and Regions. With service-managed permissions (Organizations), target OUs and auto-deploy to accounts added later; can run from a delegated admin
CloudFormation Guard (cfn-guard)Policy-as-code: write rules ("every bucket must have encryption and block public access") and validate templates (or any JSON/YAML) in the pipeline
cfn-lintChecks templates against the CloudFormation spec and best practices (syntax, invalid values)
CloudFormation HooksServer-side checks that block noncompliant resources at deploy time (proactive controls)
Stack policies, termination protection, drift detectionProtect critical resources from updates/deletes; find manual changes
CloudFormation service roleStack runs with a specific role so users don't need broad permissions themselves
Third-party IaC (Terraform, CDK)Same principles: scanning in CI, least-privilege deploy roles (OIDC from the CI system), state files encrypted and access-controlled
cfn-lint checks the grammar. Guard checks the rules of the house. Hooks are the bouncer at deploy time.

6.2Tags, Firewall Manager and resource sharing

Tags for management (6.2.2)

  • Standard keys: department, cost center, environment, owner, data classification, application.
  • Enforce with tag policies (standardize values), SCPs requiring tags on create (aws:RequestTag, aws:TagKeys), Config required-tags rule.
  • Use tags for ABAC, Resource Groups, patch targeting, backup selection, cost allocation and Firewall Manager scope.

AWS Firewall Manager (6.2.3)

Centrally deploy and enforce firewall rules across accounts. Prerequisites: Organizations, a Firewall Manager administrator account, and AWS Config enabled in the accounts and Regions.

Policy typeEnforces
AWS WAFWeb ACLs / rule groups on every ALB, API Gateway or CloudFront in scope
Shield AdvancedProtections on in-scope resources
Security groupCommon baseline SGs, audit rules (flag/fix overly permissive SGs), usage audit (unused/redundant SGs)
Network FirewallFirewalls and policies in every VPC
Route 53 Resolver DNS FirewallRule groups on every VPC
Network ACLBaseline NACL rules
Third-party firewallsPalo Alto Networks Cloud NGFW, Fortinet FortiGate CNF

Scope by OU, account or tags; new matching resources and accounts get the policy automatically, with optional auto-remediation.

Sharing resources securely (6.2.4)

ServiceShares
AWS RAMThe resource itself, no copies: VPC subnets, Transit Gateways, Resolver rules, Private CAs, prefix lists, Network Firewall policies, Backup logically air-gapped vaults and more. Enable sharing with the organization so no invitations are needed
AWS Service CatalogPortfolios of pre-approved products (CloudFormation templates) shared with accounts/OUs. Launch constraints let users deploy with a specified role, so they don't need the underlying permissions; template constraints limit parameters
"Teams must self-serve only approved, hardened architectures without being granted broad IAM permissions" → Service Catalog with launch constraints. "Every account must have the same WAF rules on every ALB, including new ones" → Firewall Manager.

6.3Evaluating compliance and remediating

Config rule detects noncompliance→Aggregator / Security Hub shows it centrally→Automatic remediation (SSM Automation)→Notification (EventBridge → SNS / User Notifications)
  • Config rules run on change or periodically; managed, custom Lambda or Guard-based; proactive mode evaluates a resource configuration before deployment.
  • Remediation: manual or automatic, using SSM Automation documents (with retries).
  • Conformance packs deploy rule sets org-wide; aggregators show compliance across accounts and Regions.
  • Security Hub CSPM standards give a score per standard and feed findings into automation.
  • Config rule compliance changes are EventBridge events: route them to SNS, tickets or chat.

6.3Collecting audit evidence

NeedAnswer
AWS's own compliance reports (SOC 1/2/3, ISO, PCI DSS attestation) and agreements (HIPAA BAA)AWS Artifact (self-service, in the console; also third-party ISV reports)
Continuously collect evidence from your accounts mapped to a frameworkAWS Audit Manager Maintenance: frameworks, assessments, evidence from CloudTrail, Config and Security Hub, assessment reports. Closed to new customers April 30, 2026; AWS points to Config conformance packs
Show current resource complianceConfig conformance pack dashboards, Config resource compliance dashboard, Security Hub CSPM standards
Prove who did whatCloudTrail with log file integrity validation, retained in the Log Archive account
Prove backups meet policyAWS Backup Audit Manager (a feature of AWS Backup, not Audit Manager)
Artifact = AWS's report card. Audit Manager = your report card.

6.3Shared responsibility and the Well-Architected security pillar

Shared responsibility model

AWS: security of the cloudYou: security in the cloud
Physical data centers, hardware, network, hypervisor, managed-service infrastructureData, IAM, encryption choices, OS patching on EC2, security groups, application code, configuration of services

The split moves with the service: on EC2 you patch the OS; on Lambda, RDS or Fargate AWS does. You always own your data, identities and configuration.

Security pillar design principles

  1. Implement a strong identity foundation
  2. Maintain traceability
  3. Apply security at all layers
  4. Automate security best practices
  5. Protect data in transit and at rest
  6. Keep people away from data
  7. Prepare for security events

AWS Well-Architected Tool: review workloads against the framework and lenses (including custom lenses for your own standards), record risks, track improvement plans and milestones.

Sources
Compare

Service families

Questions usually hinge on choosing between close neighbors. Each table lists the members, what each does, and the clue that points to it. Click any name to open its card.

Finding problems

ServiceFindsClue in the question
GuardDutyActive threats and misuse"crypto-mining", "known malicious IP", "unusual API calls", "credential exfiltration"
InspectorSoftware vulnerabilities, network exposure, code flaws"CVE", "unpatched", "container image vulnerabilities", "SBOM"
MacieSensitive data in S3"PII", "credit card numbers in buckets", "data classification"
Security Hub / CSPMPosture against standards; correlated exposures"CIS benchmark", "PCI DSS controls", "single pane of glass", "security score"
DetectiveRoot cause and scope"investigate", "root cause", "visualize", "related findings"
ConfigConfiguration changes and compliance"configuration history", "who changed", "noncompliant resources"
Access AnalyzerExternal, internal and unused access"shared outside the organization", "unused permissions", "least privilege policy"
Network Access AnalyzerUnintended network paths"no path from the internet to the database", "verify segmentation"

Logs and where they live

ServiceRecordsClue
CloudTrailAPI calls (who did what)"audit API activity", "who deleted", "organization trail"
VPC Flow LogsIP traffic metadata"rejected connections", "which IPs talked to the instance"
Resolver query logsDNS queries from VPCs"domain lookups", "DNS exfiltration"
CloudWatch LogsApp/OS/service logs, metric filters, alarms"alert when", "metric filter", "Logs Insights"
Security LakeNormalized (OCSF) security data lake"OCSF", "SIEM subscriber", "normalize across sources"
CloudTrail LakeSQL over event data stores (maintenance)"SQL queries on CloudTrail events", "7-year immutable store"
AthenaSQL over logs in S3"ad hoc SQL over logs", "serverless query"
OpenSearchSearch, dashboards, SIEM-like correlation"full-text search", "dashboards", "correlate"

Network and edge filters

ControlLayer / scopeClue
Security groupL3/L4, per ENI, stateful, allow only"only the app tier may reach the DB", "reference another SG"
Network ACLL3/L4, per subnet, stateless, allow + deny"block a specific IP", "subnet-level deny", "ephemeral ports"
Network FirewallL3–L7 for VPC traffic, IPS, domain filtering, TLS inspection"egress filtering by domain", "Suricata", "inspect traffic between VPCs"
DNS FirewallDNS queries"block malicious domains at DNS", "DNS exfiltration"
WAFL7 HTTP(S) on CloudFront, ALB, API Gateway…"SQL injection", "XSS", "rate limit", "bots", "geo block"
Shield AdvancedDDoS L3/L4 (+ L7 with WAF)"DDoS response team", "cost protection"
Firewall ManagerCentral policy for all of the above"across all accounts", "new resources automatically"

Remote and hybrid access

OptionJobClue
Session ManagerShell/port-forward with no inbound ports"no SSH keys", "no bastion", "log every command"
EC2 Instance Connect (+ Endpoint)IAM-controlled SSH; Endpoint reaches private subnets"must use SSH", "private instance without bastion"
Client VPNRemote users onto the network"remote employees need network access", "mutual certificate"
Verified AccessZero-trust per-app access, no VPN"device posture", "replace VPN", "per-request policy"
Site-to-Site VPNEncrypted site link over the internet"quickly", "IPsec", "backup for DX"
Direct Connect + MACsecPrivate link; L2 encryption on dedicated connections"consistent bandwidth", "encrypt at line rate"

Keys, secrets and certificates

ServiceJobClue
KMSManaged encryption keys integrated with services"encrypt at rest", "key policy", "rotation", "envelope encryption"
CloudHSMSingle-tenant HSMs you control"FIPS 140-3 Level 3", "PKCS#11", "sole control", "TDE"
External key storeKMS keys backed by your HSM outside AWS"keys must never be stored in AWS"
Secrets ManagerSecrets with rotation"rotate database credentials automatically"
Parameter StoreConfig and simple secrets"cheapest", "no rotation needed"
ACMPublic TLS certificates, auto-renewed"HTTPS on ALB/CloudFront", "free certificates"
Private CAPrivate certificate authority"internal TLS", "mTLS", "device certificates", "Roles Anywhere trust anchor"

Identity

ServiceForClue
Identity CenterWorkforce SSO to accounts and apps"corporate IdP", "many accounts", "permission sets", "SCIM"
CognitoCustomer identities for apps"mobile app sign-up", "social login", "JWT"
Verified PermissionsIn-app authorization in Cedar"fine-grained app permissions", "policy store"
Directory ServiceActive Directory in or connected to AWS"domain join", "trust with on-premises AD"
IAM Roles AnywhereNon-AWS workloads get temporary credentials"on-premises servers", "X.509 certificates", "no long-term keys"
STSTemporary credentials"AssumeRole", "cross-account", "external ID"

Preventive vs detective vs proactive controls

TypeExamplesEffect
PreventiveSCPs · RCPs · Declarative policies · permissions boundaries · S3 Block Public Access · VPC Block Public AccessThe action can't happen
ProactiveCloudFormation Hooks (Control Tower proactive controls) · CloudFormation Guard in CI · Access Analyzer custom policy checksBad configuration is rejected before deployment
DetectiveConfig rules · Security Hub CSPM · GuardDuty · Access AnalyzerYou find out after the fact
ResponsiveConfig remediation · EventBridge + SSM Automation / Lambda · Security Hub automation rulesIt gets fixed automatically

Immutability and recovery

FeatureProtectsClue
S3 Object Lock (compliance)S3 object versions, even from root"WORM", "regulatory retention", "SEC 17a-4"
S3 Object Lock (governance)S3 object versions, admins can override"prevent accidental deletion", "allow override"
Backup Vault LockRecovery points in a vault"backups can't be deleted early"
Logically air-gapped vaultBackups restorable from another account"recover if the account is compromised", "ransomware"
Versioning + MFA DeleteOld object versions"recover overwritten objects"
Data Lifecycle ManagerEBS snapshot schedules and copies"automate EBS snapshots"

AI security parts

FeatureJobClue
Bedrock GuardrailsPrompt attacks, PII, denied topics, grounding, Automated Reasoning"block jailbreaks", "mask PII in responses", "hallucinations"
Bedrock model invocation loggingRecord prompts and responses"audit what users asked the model"
bedrock:GuardrailIdentifier conditionForce every call to use the approved guardrail"enforce centrally"
SageMaker AI network isolationNo outbound network from training/inference containers"model must not call the internet"
Amazon Q BusinessEnterprise assistant that respects source ACLs"employees see only documents they're allowed to"
Amazon Q DeveloperSecurity scanning in the IDE"find vulnerabilities while coding"
AI services opt-out policyStop AWS AI services using your content"content must not improve AWS models"
Reference

Services A–Z

Every service and major feature you're likely to see, including the ones used as wrong answers. Each card says what it is, when to pick it, and what it gets confused with. Tags follow the SCS-C03 lists: In scope (on the list, or a security feature of a listed service), Out of scope, Not listed, plus Maintenance and Retired for services AWS changed after the guide was written.
Try it

Service flashcards

Read the situation, name the service in your head, then reveal.

Night before

Quick review: "If you see this, pick that"

Each row maps a common exam clue to the answer it usually points to.

Detection

Incident response

Infrastructure security

Identity and access management

Data protection

Governance

Acronyms A–Z

Every acronym used in this guide. Type any of them in the search box to jump here.