Cloud security engineer interview questions
The questions below come from the technical and behavioural rounds our students report after interviewing for cloud and AI platform security roles. Each one includes the answer structure panels are listening for — not a script to memorise.
AWS security questions
AWS panels lean on identity evaluation logic and multi-account design. Answer with the policy evaluation order, then the blast-radius consequence.
Walk me through how AWS evaluates a request when an SCP, a permission boundary, and an identity policy all apply.
Explicit deny anywhere wins. Otherwise the request must be allowed by the identity policy, permitted by the permission boundary, permitted by every SCP in the organizational path, and allowed by any resource policy. Name the order, then say what you would check first in a real incident: CloudTrail for the denied call, then IAM Access Analyzer for the effective permission set.
An S3 bucket holding customer exports was found publicly readable. What do you do in the first hour?
Contain first: apply the account-level public access block, then the bucket-level block, then remove the offending policy statement. Establish exposure from S3 server access logs and CloudTrail data events. Only then move to root cause — usually a policy applied by a pipeline role with no guardrail — and close it with an SCP plus an AWS Config rule.
How would you detect credential compromise for a long-lived IAM user?
GuardDuty findings for anomalous API calls and impossible-travel patterns, CloudTrail queries in Athena for calls from new regions or user agents, and access key age reporting. Then argue for removing the long-lived key entirely in favour of IAM Roles Anywhere or OIDC federation from the CI provider.
Explain KMS envelope encryption and when a customer-managed key is worth the operational cost.
Data is encrypted with a data key; the data key is encrypted by the KMS key. A customer-managed key earns its cost when you need key policies as an independent authorization layer, cross-account grants, per-tenant key separation, or a documented ability to revoke access by disabling the key.
Azure security questions
Azure interviews concentrate on Entra ID, Conditional Access, and Defender/Sentinel operations. Expect KQL on a whiteboard.
Design Conditional Access for a workforce with contractors and privileged admins.
Baseline policy requiring phishing-resistant MFA for all users, a stricter policy for privileged roles requiring compliant devices and named locations, a contractor policy scoped by group with session controls and sign-in frequency, and break-glass accounts excluded from every policy but monitored with an alert on any sign-in.
What is the difference between PIM eligible and active assignments, and why does it matter?
Eligible assignments require activation with justification, approval, and a time limit, so standing privilege drops to near zero. Active assignments are permanent. The security argument is blast radius: a compromised account with only eligible assignments cannot use the role without passing the activation controls you also alert on.
Write a KQL query to find successful sign-ins after multiple failures from the same IP.
Summarize SigninLogs by IPAddress and UserPrincipalName over a time window, count failures where ResultType != 0, join against successes, and filter for IPs above a failure threshold that later produced a success. Say out loud that you would tune the threshold against your own baseline rather than defaulting to a number.
How do you secure a storage account holding regulated data?
Disable public blob access and shared key authorization, require Entra ID authorization with RBAC scoped to containers, enforce private endpoints with a network deny rule, enable infrastructure encryption with a Key Vault key, and turn on diagnostic logging plus immutable versioning for the audit trail.
Google Cloud security questions
GCP panels test organization policy design and the IAM inheritance model more than tooling trivia.
How does IAM inheritance work across the GCP resource hierarchy?
Policies set at organization, folder, and project level are additive and inherit downward; there is no deny by default at a lower level unless you use IAM deny policies. Guardrails therefore live in organization policy constraints, not in hoping nobody grants a broad role lower down.
Which organization policy constraints would you set on day one?
Disable service account key creation, restrict public IP on Compute instances, enforce uniform bucket-level access, restrict allowed domains for IAM sharing, and require Shielded VMs. Each one removes a class of mistake rather than a single misconfiguration.
What replaces service account keys for workloads outside GCP?
Workload Identity Federation: exchange an external OIDC or SAML token for short-lived credentials, with attribute conditions restricting which repository, branch, or cluster may assume the service account.
AI platform security questions
Increasingly common in 2026. Panels want to know you can reason about model-facing systems, not recite OWASP lists.
How do you threat model a retrieval-augmented application?
Treat retrieved content as untrusted input to the model. Controls: strict tool allow-lists, no privileged actions triggered by model output without a deterministic authorization check, tenant-scoped retrieval so a query cannot reach another customer's documents, and output logging for abuse review.
What stops prompt injection from escalating into data loss?
Authorization outside the model. The model may propose an action; the server verifies the caller's identity and permissions before executing it. Add egress restrictions so an injected instruction cannot exfiltrate to an arbitrary host.
Behavioural and scenario rounds
Use situation, action, result — and make the result a measurable security outcome, not a feeling.
Tell me about a time you disagreed with an engineering team about a security requirement.
Describe the risk in their language (downtime, customer impact, audit finding), the compromise you proposed, and what shipped. Panels are testing whether you can be overruled gracefully and still reduce risk.
Describe an incident you handled end to end.
Detection source, containment decision and its trade-off, evidence you preserved, root cause, and the control you added so the same class of incident cannot recur. Name what you got wrong; interviewers trust candidates who do.
You have no cloud security job title yet. Why should we hire you?
Point at artefacts: environments you built and broke, detections you wrote, write-ups you published. Specific evidence beats a certification list in every panel we have seen.
How to prepare in the two weeks before an interview
- Rebuild one broken environment per platform and be able to narrate the fix from memory.
- Practise policy evaluation out loud — most candidates know the rules but cannot sequence them under pressure.
- Bring three incident stories with measurable outcomes; you will reuse them in every behavioural round.
- Prepare two questions for the panel about their detection coverage and on-call model.
Build the evidence these answers need
Our 20-week program is built around 36 projects plus a capstone, so every answer above maps to something you actually built.