Based on the SCS-C03 exam guideReport an error
This is the last day of the 12 weeks. Today we organize not new knowledge but the skill of turning it into points in the exam room. SCS-C03 has 65 questions in 170 minutes, and the passing score is 750/1000 (scaled). The final gate is moving what you know into answers on time and avoiding the traps.
Specialty is not a "knowledge exam" but a "judgment exam." Even if both options work, the one that better fits the principles of the AWS Well-Architected security pillar (least privilege, defense in depth, traceability, automated security, protecting data in transit and at rest) is the answer. A last-minute two-way choice almost always splits on one of these principles.
The one word at the end of the question is the grading criterion. Skim this condensed version of yesterday's table once just before the exam.
| Qualifier | Direction the answer leans | Representative candidates |
|---|---|---|
| MOST cost-effective | Free, built-in, fewer calls | Gateway Endpoint, Parameter Store, S3 Bucket Keys |
| LEAST operational overhead | Managed > self-built | Secrets Manager rotation, managed rules, Control Tower |
| MOST secure | Narrower permissions, stronger immutability | Object Lock Compliance, SSE-KMS, PrivateLink |
| FASTEST to implement | Existing managed combinations | Control Tower, Conformance Pack |
| MINIMAL disruption | Non-invasive, staged | WAF count mode → block, warn then enforce |
| without changing the application | Solve at the infrastructure layer | Edge, endpoints, agents |
| automatically / immediately | Event-driven | EventBridge → Lambda/SSM |
| prevent | Make the action itself impossible | SCP Deny, proactive controls |
| all / future accounts | Organization tools | SCP/RCP, Firewall Manager, auto-enable |
The reflex of instantly translating the phrasing in the question into a service. This is the key to speed. We've split it by area.
| When the question says this | Service/pattern that comes to mind |
|---|---|
| "Anomalous behavior, malicious IP, suspected data exfiltration, crypto mining" | GuardDuty |
| "Malware on an EBS volume" | GuardDuty Malware Protection |
| "Process-level behavior of containers and nodes" | GuardDuty Runtime Monitoring |
| "Root cause, scope of compromise, lateral movement, visual investigation" | Detective |
| "Vulnerabilities, CVEs, missing patches, network reachability" | Inspector |
| "Discover and classify PII, card numbers, and sensitive data in S3" | Macie |
| "Find policies exposed externally or cross-account" | IAM Access Analyzer (external access) |
| "Clean up unused permissions and roles" | IAM Access Analyzer (unused access) |
| "Finding aggregation, CIS/PCI/FSBP scores" | Security Hub |
| "Raw logs to a third-party SIEM" | Security Lake (OCSF) |
| "Too many findings, noise" | Security Hub automation rules + Insights |
| "Findings from multiple Regions on one screen" | Security Hub aggregation Region |
| "New accounts are also automatically in scope for detection" | Delegated administrator + auto-enable |
| When the question says this | Service/pattern that comes to mind |
|---|---|
| "Who called which API" | CloudTrail management events |
| "Which objects were downloaded or uploaded" | CloudTrail data events (enabled separately) |
| "Call volume or error rate surges above normal" | CloudTrail Insights |
| "Centralize activity from all accounts" | Organization trail |
| "Prove the logs have not been tampered with" | log file validation (SHA-256 digest) |
| "Prevent deleting or turning off" | SCP Deny + Log Archive account + Object Lock |
| "What state was this resource in at that time" | Config configuration item timeline |
| "Configuration compliance and drift evaluation" | AWS Config rules |
| "Multi-account, multi-Region Config data in one place" | Config aggregator |
| "IP flows, accept/reject only" | VPC Flow Logs |
| "Need to see even the packet contents" | Traffic Mirroring |
| "Which domains were queried" | Route 53 Resolver query log |
| "Immediate alert on root login" | CloudTrail → CloudWatch Logs metric filter → alarm → SNS |
| "Investigate large volumes of logs with SQL" | Athena / CloudTrail Lake |
| When the question says this | Service/pattern that comes to mind |
|---|---|
| "S3/DynamoDB without going through the internet" + "minimize cost" | Gateway VPC Endpoint (free) |
| "Other services / my SaaS without going through the internet" | Interface Endpoint (PrivateLink) |
| "Access the bucket only from this VPC" | Bucket policy aws:SourceVpce / aws:SourceVpc |
| "Block a specific IP range (subnet)" | NACL deny |
| "Per-instance allow list" | Security Group |
| "Inbound is open but no response comes back" | NACL outbound ephemeral ports (1024–65535) |
| "Must cut an in-progress connection immediately" | NACL (SGs are stateful) |
| "L7 SQLi/XSS/rate limit/geo blocking" | WAF |
| "Absorb L3/4 DDoS, cost protection, response team" | Shield Advanced |
| "VPC IPS, domain control, central inspection" | Network Firewall (+ TGW appliance mode) |
| "DNS-based malware/exfiltration" | Route 53 Resolver DNS Firewall |
| "Prevent direct access to the origin" | CloudFront + OAC / secret header / prefix list |
| "Hub for many VPCs, transitive routing" | Transit Gateway (Peering is non-transitive) |
| "Dedicated on-premises bandwidth" | Direct Connect (+ VPN encryption) |
| "Third-party appliance inline" | Gateway Load Balancer |
| When the question says this | Service/pattern that comes to mind |
|---|---|
| "Organization permission ceiling, prevent doing ~" | SCP |
| "Block external access to organization resources" | RCP |
| "Delegated person must not exceed their own permissions" | Permissions boundary |
| "Safe third-party delegation, confused deputy" | AssumeRole + trust policy + External ID |
| "IAM role for on-premises workloads" | IAM Roles Anywhere (X.509) |
| "Multi-account SSO, SAML/OIDC federation" | IAM Identity Center |
| "CI/CD without long-term keys" | OIDC federated role |
| "App end-user authentication" | Cognito (User Pool) / temporary AWS credentials (Identity Pool) |
| "Credentials for EC2/EKS/ECS/Lambda" | Instance profile / IRSA, Pod Identity / task role / execution role |
| "Prevent metadata theft via SSRF" | IMDSv2 + hop limit |
| "Automatic rotation of DB credentials" | Secrets Manager |
| "Settings/secrets (no rotation needed, low cost)" | SSM Parameter Store (SecureString) |
| "Invalidate immediately" | Access key Inactive / aws:TokenIssueTime Deny |
| When the question says this | Service/pattern that comes to mind |
|---|---|
| "We solely own the keys, dedicated HSM, FIPS L3" | CloudHSM / KMS custom key store |
| "With keys from an external key management system" | KMS external key store (XKS) |
| "Key usage auditing, policy-controlled encryption" | SSE-KMS + CloudTrail |
| "Dual-layer encryption" | DSSE-KMS |
| "Keys are not kept at AWS" | SSE-C / client-side encryption |
| "It's SSE-KMS but KMS cost is excessive" | S3 Bucket Keys |
| "Use the key only when it goes through a specific service" | kms:ViaService |
| "Separate decryption per tenant" | kms:EncryptionContext |
| "No deletion or tampering, WORM, ransomware prevention" | S3 Object Lock + MFA Delete + versioning |
| "No one (including root) can delete for N years" | Object Lock Compliance |
| "Prevent mistakes, allow exceptions when needed" | Object Lock Governance |
| "Backups must not be deleted" | AWS Backup Vault Lock |
| "Reject plaintext HTTP" | Bucket policy aws:SecureTransport Deny |
| "Automatic certificate renewal" | ACM (CloudFront is us-east-1, imported ones don't renew) |
| "Internal chain of trust" | AWS Private CA |
| "Encrypt an EBS/RDS that already exists" | Snapshot → encrypted copy → restore |
| "Volumes created from now on are automatically encrypted" | Account/Region EBS default encryption |
| "Centrally enforce organization WAF/SG/Shield" | Firewall Manager |
| "Automatic guardrails for new accounts, landing zone" | Control Tower |
| "Block in IaC before creation" | Proactive controls (CloudFormation Hooks) |
| "Roll out a set of rules to the organization at once" | Conformance Pack |
| "Evidence and reports for submission to auditors" | Audit Manager |
| "Self-service only for approved products" | Service Catalog |
| "Share resources cross-account" | RAM |
| "Standardize tag notation" | Tag Policy |
| "Block creation without tags" | SCP (aws:RequestTag + Null) |
| "A spending spike is a sign of compromise" | Cost Anomaly Detection / Budgets |
| "Data must not leave the organization" | SCP (outbound door) + RCP (inbound door) + aws:PrincipalOrgID |
Most lost exam points come not from not knowing but from blurry boundaries. You can reread just this section right before the exam.
| Service | Question it answers | One word |
|---|---|---|
| GuardDuty | "Is a threat behavior happening now?" | Threat |
| Inspector | "Is an exploitable vulnerability left?" | Weakness |
| Macie | "What data is sensitive and where is it?" | Data |
| Detective | "How are these entities connected?" | Relationships |
| Access Analyzer | "What is open to the outside?" | Exposure |
| Security Hub | "How do I see all of these in one form?" | Aggregation |
| Policy | Attaches to | Grants permission | Representative use |
|---|---|---|---|
| SCP | Root/OU/account | ✗ | Organization ceiling: Region lock, protecting security services. Does not apply to the management account |
| RCP | Root/OU/account | ✗ | Resource-side organization limit: blocks external principals. Limited supported services |
| Permissions boundary | Individual user/role | ✗ | The ceiling of a single principal: keeps a delegator from exceeding permissions |
| Resource policy | S3, KMS, SQS, Lambda, etc. | ○ | Cross-account sharing, allowing service principals |
| Session policy | AssumeRole session | ✗ | Temporary narrowing of temporary credentials |
One-line rule: If the scope is an account or OU, SCP; if it is a single principal, permissions boundary; if it is the resource side, resource policy or RCP.
| Axis | Key policy | IAM policy | Grant |
|---|---|---|---|
| Sufficient alone | ○ | ✗ (premised on the key policy's delegation) | ○ (within the allowed scope) |
| Lifetime | Permanent | Permanent | Temporary, revocable |
| Typical use | Key ownership boundary | Usage permissions for people and apps | A service using the key on the user's behalf |
The answer to "I have AdministratorAccess but can't decrypt" is usually the key policy.
| Axis | Governance | Compliance | Legal Hold |
|---|---|---|---|
| Deletion during retention | Possible with special permission | Not possible (including root) | Not possible until released |
| Period | Specified, can be shortened | Specified, cannot be shortened (only extended) | No period |
| When | Internal policy, preventing mistakes | Regulatory requirement | Freezing for litigation or investigation |
| Axis | SSE-S3 | SSE-KMS | DSSE-KMS | SSE-C |
|---|---|---|---|---|
| Key control and auditing | Weak | Key policy + CloudTrail | Same | Entirely the customer |
| Encryption layers | 1 | 1 | 2 (dual) | 1 |
| Clue for choosing | Default encryption | "Auditing, control" | "Dual layer" | "Keys not kept at AWS" |
| Axis | Config | CloudTrail | Flow Logs |
|---|---|---|---|
| What it answers | "What state was it in then?" | "Who called what?" | "Where did it flow?" |
| Unit | Resource configuration item | API event | ENI/subnet/VPC flow |
| Payload | — | — | None (metadata only) |
| Role in an investigation | When it changed | Who did it | Where it went out to |
| Threat or requirement | Answer |
|---|---|
| L3/4 volumetric flood, DDoS cost protection | Shield Advanced |
| L7 SQLi/XSS/bots/rate limit/geo | WAF |
| IPS, domains, and signatures inside the VPC and for egress | Network Firewall (forced routing required) |
| Subnet-level IP blacklist | NACL deny |
| Per-instance allow list | Security Group |
Finally, we gather the most expensive mistakes in one place.
⚠️ Logging and Detection Traps:
- CloudTrail management events do not record S3 object and Lambda data access → enable data events separately.
- GuardDuty consumes logs internally, so there is no need to turn on Flow Logs/DNS logs separately for it; don't turn them off.
- GuardDuty = threats, Inspector = vulnerabilities, Macie = data, Detective = investigation, Security Hub = aggregation. Don't confuse them.
- VPC Flow Logs don't see the payload (accept/reject and metadata only). For the contents, use packet mirroring or app logs.
⚠️ IAM and Governance Traps:
- SCPs and Permission Boundaries do not grant permissions; they only limit the ceiling.
- Evaluation order: an explicit Deny takes top priority, then SCP, Allow, Boundary, implicit Deny.
- Don't put long-term access keys on workloads → roles (instance profile/IRSA/execution role).
- Prevent confused deputy in cross-account delegation with an External ID.
- IAM roles = employees and workloads, Cognito = app users.
Network Traps
Data Protection Traps
aws:SecureTransport) are separate controls: both are needed.Incident Response Traps
accessKeyId lookup axis.Governance and Proof Traps
RemediationEnabled is off, it only reports.🎯 Last-minute two-way choice decision rule: If both answers work: (1) Is it more managed? (2) Is it more automated? (3) Is it more least-privilege? (4) Is it more non-bypassable? (5) Is it earlier in the prevent > detect > respond hierarchy? (6) Can it be enforced organization-wide? (7) Is it reversible (for automated-response questions)? The one that satisfies more of these is the best.
Both options "work"
│
├─① Is there a qualifier? ── If so, that axis is the grading criterion. It ends here.
│ MOST cost-effective → the free/built-in side
│ LEAST operational overhead → the managed side
│ MOST secure → the narrower and more immutable side
│
├─② Does one side violate a constraint?
│ "no internet" / "existing app" / "multi-account" / "regulatory"
│ → if even one is violated, that option is eliminated immediately
│
├─③ Does the requested verb match the control layer?
│ prevent → SCP, policy Deny, proactive controls
│ detect → Config, GuardDuty, Access Analyzer
│ respond → EventBridge, Lambda, SSM
│ (an option that says it will block using a detection tool is wrong)
│
├─④ Is it a half measure?
│ Locks only the path while leaving permissions as is / only containment with no tracing or eradication
│ → the side with the complete combination is best
│
├─⑤ Is it an extreme statement?
│ "automate everything" "everything manual" "turn off detection" "revoke all permissions"
│ → extremes are usually wrong
│
└─⑥ If it still splits ── the preference hierarchy
managed > self-managed
automated > manual
prevent > detect > respond
organization-enforced > per-account
reversible automation > irreversible automation
The knowledge is already done. The remaining variables are condition and procedure.
The day before — don't open new material. Skim only the comparison tables and trap list of today's document (new knowledge only adds anxiety). Check the exam location, time, and required ID, and if taking it online, finish the system check and tidy your desk in advance. Sleep enough: 170 minutes of concentration is decided in the last 20 questions, and accuracy in that stretch is directly proportional to sleep.
On the day — right at the start, spend 30 seconds to re-engrave the checkpoint numbers in your head (85 minutes at question 33). Pass 1 goes quickly, centered on questions you are confident about, and for flagged questions always fill in a tentative answer before moving on. If you check the on-screen flag-for-review feature and the remaining time display early, you won't wander in the middle. If you feel your pace has collapsed midway, return to the order of reading the stem: last sentence → constraints → options.
The exam looks less at "do you know this service" than at "do you pick the right combination in this situation." So the flow of scenario decomposition → keyword translation → trap avoidance → applying the preference hierarchy makes the score more than memorizing single answers. The decisiveness to flag and move past a question you don't know protects the overall score. The pace of 65 questions, not the perfection of one question, makes the pass. Just apply the per-domain mental models built over 12 weeks as they are: the detection and response nervous system (1·2), the dual control of path × permissions (3·4), and encryption × governance propagation (5·6). And one more thing: nervousness makes you read the stem shortly. The most common way to miss a question you know in the real exam is "skipping one line of constraints." The place to speed up is eliminating options, not reading the stem.
If you recall in one line each which incidents the judgment this certification demands actually came from, the trap list stays as stories rather than memorization. Code Spaces (2014): with the account taken over, the data and even the backups in the same account were deleted, and the company shut down. → Put logs and backups in an account that the creator cannot delete. Capital One (2019): SSRF reached the instance metadata and obtained the role's temporary credentials, data was exfiltrated, and awareness came from an external tip. → Isolation and credential revocation are separate, IMDSv2 and least privilege are preparation items and not response, and without your own detection loop even awareness comes late. Equifax (2017): Entry came through a missed patch, and the certificate of the traffic inspection device was left expired, so detection did not work for a long period. → A control must be confirmed "working now," not merely "installed." SolarWinds (2020): It infiltrated through a trusted supply chain and was finally revealed by one security company's own anomaly detection. → Trust boundaries must be drawn on the inside too, and the final line of defense is always your own detection. The reasons that "a person checks regularly" is a wrong answer in the exam options and that "store immutably in a separate account" is the right answer are all within these four sentences.
Compressing the 12 weeks into one sentence: "Logs make the evidence, detection makes the judgment, boundaries and permissions limit the damage, and encryption and governance keep that state in place."
There are only three things you actually use in the exam room.
Once you have these three, the rest is a matter of pace. 85 minutes at question 33, 150 minutes for pass 1, 20 minutes on flagged questions. Leave no question blank.
One last thing. Meeting a question you don't know is not a sign of failure but the normal distribution. The passing score is not a perfect score, and some questions mixed in are not even scored. The real danger is being shaken by one question and losing your focus on the next five. Flag it, move on, and come back.
Just decompose calmly, translate, and avoid the traps.
Click a choice to reveal the answer and explanation.
Question 1
During the exam, you have been unable to decide between two options on one question for 4 minutes. What is the most reasonable action?
Question 2
The question says "automatically correct non-compliant resources without human intervention (automatically, without manual intervention)." What typical pattern does this keyword point to?
Question 3
The requirement is "block a specific action so it cannot be done in any account of the organization." The two options are (A) detect with a Config rule and (B) Deny with an SCP. Which is best?
Question 4
Both options work: (A) a script on EC2 that rotates keys directly with cron, and (B) Secrets Manager automatic rotation. What is the preferred answer in SCS-C03 and why?
Question 5
Which of the following trap statements is not true?
Question 6
The requirement is "access S3 from a private subnet in the MOST cost-effective way." What answer does the emphasis word "MOST cost-effective" decide?
Question 7
Which is the most inappropriate exam mindset?
Question 8
Which pairing connects the two requirements, "automatically discover whether credit card numbers are stored in S3" and "assess whether EC2 has unpatched CVEs," to the correct services?