SAA-C03 domain - 30% of the exam

Design Secure Architectures

Design Secure Architectures is 30% of the SAA-C03 exam. These are the objectives it covers, each with practice questions, with every answer explained.

The domain in numbers

109
Practice questions
9
Objectives
83
Medium
26
Hard

Where people go wrong

  • Using permissions boundaries where an SCP is needed

    The wrong answers reach for a permissions boundary when the requirement is a guarantee across every account in an organizational unit. A boundary attaches to individual principals, so a role created without it escapes the control, and Identity Center recreates provisioned roles. A service control policy on the OU caps every principal in the member accounts. The reverse also appears: an SCP cannot single out developer-created roles, which is what a boundary is for.

    See it in a sample question
  • Granting cross-account access from one side only

    The wrong answers grant a cross-account read from one side only: an identity policy in the consumer account alone, a bucket policy that ignores the separate decrypt step, or an AWS managed key, which cannot be shared across accounts. Cross-account KMS use needs the key policy in the owner account to allow the external account and an identity policy in the consumer to permit kms:Decrypt. A secret in another account also needs a resource policy on the secret, plus kms:Decrypt on its customer managed key.

    See it in a sample question
  • Confusing the AWS security detection services

    Each wrong option names a service that sounds like a detective control but covers a different job. Macie classifies sensitive data inside S3 objects. Inspector assesses EC2, ECR images and Lambda for known CVEs and unintended network reachability. GuardDuty analyses VPC Flow Logs, DNS logs and CloudTrail for malicious behaviour. Security Hub aggregates findings from those services and runs standards checks, but originates no log-based detections itself.

    See it in a sample question
  • Mixing up security groups and network ACLs

    The wrong answers attach the behaviour of one control to the other. Security groups are stateful, allow rules only, and bind to the instance, so return traffic is permitted automatically and a rule can name another security group as its source. Network ACLs are stateless, bind to the subnet, evaluate numbered rules in order, and support explicit deny, so responses need their own outbound rule on the ephemeral port range.

    See it in a sample question
  • Expecting Parameter Store to rotate credentials

    The wrong answers choose Systems Manager Parameter Store, as a String or a SecureString, when the requirement is scheduled rotation of a database password. A SecureString encrypts the value but has no native rotation or cross-Region replication, so the team would build the schedule, the new password and the database update itself. Secrets Manager rotates supported RDS engines natively, runs a custom rotation Lambda for other databases, and replicates a secret across Regions.

    See it in a sample question

Objectives in this domain

What to study

The official documents this domain's practice questions cite most.

Sample questions from this domain

Free sampleDesign Secure Architectureshard

A company runs 60 accounts in AWS Organizations with all features enabled. Security wants to guarantee that no IAM principal in any member account in the Production organizational unit can disable AWS CloudTrail, regardless of the permissions an account administrator later grants. Which approach enforces this guarantee with the least ongoing effort?

  • AAttach an IAM permissions boundary that denies CloudTrail write actions to every role created in each Production member account through a deployment pipeline.
  • BAttach a service control policy to the Production organizational unit that denies cloudtrail:StopLogging and cloudtrail:DeleteTrail for all principals in those accounts. Correct
  • CEnable an organization trail and rely on the management account administrators to monitor and manually re-enable CloudTrail whenever a member account stops it.
  • DCreate a deny-based resource policy on each CloudTrail trail in the Production accounts so that StopLogging and DeleteTrail calls are rejected for member principals.
Use service control policies on an organizational unit to set an unbreakable permission ceiling across all member accounts. A service control policy applied to an organizational unit defines the maximum permissions for every principal in the affected accounts. Because effective permissions are the intersection of the SCP and any identity policy, an explicit deny in the SCP cannot be overridden by an account administrator granting the action later.

Why A is wrong: Permissions boundaries cap a principal's permissions but must be attached to each principal individually, so a new role created without the boundary would escape the control.

Why B is correct: An SCP on the OU sets a permission ceiling that no IAM policy in any member account can exceed, so the deny holds even if an administrator later grants the action.

Why C is wrong: An organization trail centralises logging but does not prevent a member principal from stopping logging, and manual remediation is reactive rather than a guarantee.

Why D is wrong: CloudTrail trails do not support a resource policy that blocks management actions, so this control cannot be applied and offers no real enforcement.

Free sampleDesign Secure Architecturesmedium

A data team in a producer account encrypts objects in an Amazon S3 bucket using a customer managed KMS key in that same producer account. A consumer account must be allowed to download and decrypt those objects through a role it already owns. The producer wants to grant only decrypt access on that one key and keep the grant defined where the key lives. Which configuration meets these requirements?

  • AAttach an identity policy to the consumer role allowing kms:Decrypt on the key Amazon Resource Name, and make no change to the KMS key policy in the producer account at all.
  • BCopy the KMS key material into the consumer account and recreate an identical key there so the consumer role can call kms:Decrypt against its own local copy of the key.
  • CEnable an Amazon S3 bucket policy that grants the consumer account s3:GetObject, which automatically authorises the consumer role to decrypt the objects with the producer's KMS key.
  • DAdd a statement to the KMS key policy in the producer account allowing kms:Decrypt for the consumer account, then have the consumer role's identity policy permit kms:Decrypt on that key. Correct
Cross-account use of a customer managed KMS key requires both a key policy grant in the owning account and a matching identity policy in the caller account. KMS evaluates the key policy of the account that owns the key and the identity policy of the calling principal. For a cross-account decrypt, the owner account's key policy must allow the external account and the caller's identity policy must permit kms:Decrypt; an S3 bucket policy does not authorise the separate KMS operation.

Why A is wrong: An identity policy alone is tempting because it works for same-account keys, but a cross-account caller is still denied unless the key policy in the owner account also grants the external account access.

Why B is wrong: KMS key material cannot be exported and copied between accounts like this, and a separately created key could not decrypt ciphertext produced by the producer's distinct key.

Why C is wrong: A bucket policy can grant object read, but decryption is a separate authorisation against the KMS key, so without a key policy grant the GetObject call still fails on the decrypt step.

Why D is correct: Cross-account KMS access needs both the key policy in the owner account to allow the external principal and an identity policy in the consumer account that permits the call, so this dual grant is correct.

Free sampleDesign Secure Architecturesmedium

A development team runs a fleet of EC2 instances, several container images in Amazon ECR, and a number of Lambda functions. Security wants an automated, managed assessment that continuously checks these workloads for known software vulnerabilities such as outdated packages with published CVEs and flags unintended network reachability, without the team building and maintaining its own scanning infrastructure. Which service meets this requirement?

  • AAmazon Macie, which continuously assesses the EC2 instances, container images, and functions for known software vulnerabilities and risky network reachability
  • BAWS Security Hub, which scans the EC2 instances, container images, and functions itself to find vulnerable packages and unintended network exposure
  • CAmazon GuardDuty, which continuously scans the EC2 instances, container images, and functions for vulnerable packages and risky network reachability
  • DAmazon Inspector, which continuously assesses EC2, ECR images, and Lambda for known CVEs and unintended network reachability Correct
Amazon Inspector is the managed service that continuously assesses EC2, ECR images, and Lambda for known software vulnerabilities and network reachability. Inspector maintains a software inventory of supported workloads and automatically compares installed packages against published CVEs while analysing network paths for unintended exposure, so it delivers continuous vulnerability assessment without the team operating its own scanning infrastructure.

Why A is wrong: Macie discovers and classifies sensitive data in S3, so it is a managed security service, but it does not scan compute workloads for vulnerable packages or evaluate network reachability against published CVEs.

Why B is wrong: Security Hub aggregates findings and runs configuration checks against standards, so it appears to cover vulnerabilities, but it relies on services such as Inspector for the actual workload scanning rather than performing CVE assessment itself.

Why C is wrong: GuardDuty detects malicious and anomalous activity from logs and is tempting because it is continuous, but it does not enumerate installed packages or match them to CVEs, so it is not a vulnerability assessment service.

Why D is correct: Inspector is the managed vulnerability assessment service that continuously scans EC2 instances, ECR container images, and Lambda functions, matching software inventory to known CVEs and evaluating network reachability without any self-managed scanner.

Free sampleDesign Secure Architecturesmedium

An architect places database EC2 instances in a private subnet and adds a security group rule allowing inbound TCP 3306 from the application tier security group. The application servers can open connections to the database, but the architect notices no explicit security group rule permitting the database to send responses back to the application servers. Why do the responses still reach the application tier?

  • ASecurity groups are stateful, so return traffic for an allowed inbound connection is permitted automatically regardless of the outbound rules. Correct
  • BNetwork ACLs on the subnet are stateful and silently create a matching return-path entry for every connection that the security group admits.
  • CThe default security group outbound rule allows all traffic, and that egress rule alone is what forwards the database responses back to the application tier.
  • DThe implied VPC router rewrites the response packets so they bypass security group evaluation entirely on their way back to the application tier.
Security groups are stateful, so allowed inbound connections automatically permit their return traffic without a matching outbound rule. A security group records the state of each connection it admits. When the database receives an allowed inbound connection on port 3306, the response packets are treated as part of that established flow and are allowed out automatically, independent of any outbound rule configuration.

Why A is correct: Security groups track connection state, so when an inbound flow is allowed the corresponding response traffic is permitted automatically without an explicit outbound rule.

Why B is wrong: This is tempting because both controls handle return traffic, but network ACLs are stateless and require explicit rules for response traffic, so they are not the reason here.

Why C is wrong: While the default egress rule does allow all outbound traffic, response delivery is driven by connection state tracking, not by a separate egress rule evaluating the reply.

Why D is wrong: The VPC router forwards packets between subnets but never bypasses security group evaluation; it does not rewrite or exempt response traffic from filtering.

Free sampleDesign Secure Architecturesmedium

A company stores credentials for a self-managed MariaDB database on EC2 in a secrets service and must rotate them automatically every 14 days. The database is not one of the engines that Secrets Manager rotates out of the box, so rotation must run custom logic against the database. Which approach meets the rotation requirement with the least ongoing effort?

  • AKeep the credentials in Systems Manager Parameter Store as a SecureString and schedule an EventBridge rule that triggers a Lambda function to write a new value every 14 days.
  • BStore the credentials in AWS Secrets Manager and attach a custom rotation Lambda function on a 14 day schedule that updates the password in MariaDB and in the secret. Correct
  • CStore the credentials in AWS Secrets Manager and enable the built-in managed rotation option, which detects the MariaDB engine and rotates the password without any function.
  • DStore the credentials in Parameter Store as a String parameter and rely on parameter version history so an operator can revert to a previous password whenever a rotation is required.
Use Secrets Manager with a custom rotation Lambda for secrets whose target engine is not natively supported by managed rotation. Secrets Manager schedules rotation, generates the new version, manages the AWSCURRENT and AWSPENDING staging labels and calls a rotation Lambda; for an unsupported database you supply only the function that changes and tests the password, so the service still removes most of the rotation effort that Parameter Store would leave to you.

Why A is wrong: Parameter Store has no rotation feature, so you must build the schedule, the new-password generation, the database update and error handling yourself, which is more bespoke plumbing than the managed option.

Why B is correct: Secrets Manager invokes a rotation Lambda on the schedule you set and manages versioning and the staging labels, so a custom function only implements the database steps while the service handles the rest.

Why C is wrong: Managed rotation with no function applies only to natively supported targets, and a self-managed database on EC2 is not one of them, so this option would not actually perform the rotation described.

Why D is wrong: A String parameter is plaintext and version history only records edits, so this neither encrypts the secret nor performs any automatic rotation against the running database.

Other domains in this exam

See also the SAA-C03 cert hub, the study guide, and the cheat sheet.

Examworthy is not affiliated with or endorsed by Amazon Web Services. Original, blueprint-aligned practice material only.