Skip to content
Tech Interview Prep
Technical interview guide

Defense in Depth

Layering multiple independent security controls so no single failure compromises the whole system.

Read
29 min
Practice MCQs
25
Interview Q&A
25
Edition
v4
Editorial status
Reviewed

Scope: NIST SP 800-53 Rev. 5 update 1; SP 800-37 Rev. 2; SP 800-160 Vol. 2 Rev. 1; SP 800-207; SP 800-61 Rev. 3; CIS Controls v8/8.1; OWASP ASVS project current 2026-08-31; MITRE ATT&CK current 2026-08-31.

Curated: · Written: · Reviewed:

25 questions

MCQ-15A team runs a customer-data API on a managed Kubernetes service and stores PII in S3. A review finds the S3 bucket was world-readable for 6 hours after a deploy overwrote its bucket policy. The team lead writes: 'AWS owns security of the cloud, so this is AWS's gap.' Which conclusion should the review record?
MCQ-21A payments API is defended by three layers: an edge WAF, an in-service authorization check, and database row-level security. Single-layer tests show each lets through 20%, 10% and 5% of crafted attack requests respectively. A red-team run of 2,000 crafted requests produced 180 that defeated all three layers and reached the protected data. What does this result establish?
MCQ-16A payments team runs three controls in series in front of the card vault: an edge WAF, a service-level authorization check, and a vault token check. A red team exercised 80 attack paths. With all three controls live, the WAF stopped 50; of the 30 that got past it, the authorization check stopped 18; of the remaining 12, the token check stopped 9, so 3 paths reached the vault. The team then replayed all 80 paths with one control disabled at a time: with the WAF off, 22 paths reached the vault; with authorization off, 11; with the token check off, 6. The team wants one number for the board that captures the depth of the design, how well it holds up when any single control fails, rather than the strength of any one layer. Which figure should they report?
MCQ-17A payments API is fronted by two WAFs in series. Both run the same vendor's managed rule set, both are configured from one Terraform module, and both authenticate admins through the same SSO tenant. A managed-rule update rolled out at 14:02 began blocking legitimate tokenization callbacks; both WAFs started blocking within the same second, and the on-call could not disable one layer while leaving the other up. What does this incident show about the second WAF?
MCQ-1A payments service sits behind two WAF rule sets in series, which the team cites as its defense in depth. On a 1,000-payload labelled corpus of injection attacks, rule set A flags 900, rule set B flags 880, and the union flags 910. A third rule set from the same vendor and rule family is proposed. What should the team expect from adding it?
MCQ-2A payments team already runs an API gateway that validates each caller's JWT against a JWKS endpoint and enforces per-route scopes from an OPA bundle. They want a second control on the same request path. Which addition is least independent of the gateway?
MCQ-3A checkout service sits behind three controls in series: a WAF, an API gateway that validates tokens, and a service mesh enforcing mTLS. In a red-team exercise each control alone blocked about 90% of the attack variants. In series the team measured 91%, not the 99.9% they predicted. All three load their rules from the same nightly config bundle, and the variant that got through exploited a parsing bug present in that shared bundle. Which conclusion should drive the next change?
MCQ-25A team operates five independent layers (a WAF, mutual TLS, input validation, per-request authorization, and an audit log) and each one reports healthy on its own dashboard. A red-team exercise still found a WAF rule that stopped matching after a vendor update and an authorization check that returned allow on a malformed token. How should control health be monitored so a silently degraded layer is caught before an attacker uses it?
MCQ-22A SOC's endpoint EDR, network IDS, and cloud audit detections all consume events from one host-resident log forwarder; an attacker who disables that forwarder silences every alert. Which design change most directly restores detection depth?
MCQ-14A team maps 143 ATT&CK techniques to its EDR, SIEM and SOAR controls and reports 87% coverage; the dashboard counts a technique as covered when any control maps to it, without checking whether that control's telemetry reaches the hosts the technique targets. A purple-team run then emulates 20 credential-access techniques against the server tier; 14 raise no alert. Nine of the 14 have a SIEM rule that needs Sysmon Event ID 8, which is deployed on endpoints but not on servers, and server-side Sysmon is not licensed this year. Which conclusion does the run support?
MCQ-9A payments service runs as a pod on a node, writing JSON logs to a hostPath volume; a sidecar ships them to the collector using a token mounted into the pod. During an incident, an attacker with root in the container deleted the local files and used the mounted token to POST 12,000 forged "no anomaly" entries covering the 40-minute intrusion. Which change would have left the collector's record trustworthy for that window?
MCQ-4An internal REST API logs response bytes per API key at the gateway, and responses average 120 KB. Its loudest legitimate caller is a nightly reporting export that moves 3 GB per key in a 10-minute burst; ordinary keys move a few hundred MB a day. Over six days one API key pulled 40 GB out, spread evenly, and nothing fired: every request carried a valid key and was well-formed, so the WAF stayed quiet. Before the next audit the team can add one per-key rule. Which rule catches this exfiltration without alerting on the export job?
MCQ-7During a red-team exercise, an attacker moved from a compromised frontend pod to the billing database in 90 seconds. The cluster runs the default allow-all pod network, and the database accepts any caller presenting the cluster's shared service certificate. Which change most directly removes that path?
MCQ-10A 1,200-person company runs a flat /16 corporate network with no internal segmentation, so any host can reach any other. It puts an identity-aware proxy in front of its internal apps: every request is authenticated and the device's posture is checked. An attacker phishes an engineer's credentials and, using malware on that engineer's enrolled laptop, replays the stolen session cookie from the same laptop. Posture checks pass. What happens next?
MCQ-24A platform team runs 300 services across 12 namespaces on a Kubernetes cluster. The standard profile creates each namespace with a default-allow NetworkPolicy, mounts a service account token with cluster-wide read into every pod, and exposes any service annotated `public: true` through a cloud load balancer. The secure profile creates each namespace with default-deny ingress and egress, does not mount a service account token unless the pod requests one, and requires an explicit allowlist annotation for public exposure. An RCE compromises a pod in the `payments` namespace in the secure profile that did not request a token. The attacker runs `curl http://169.254.169.254/latest/meta-data/iam/security-credentials/` and then tries to connect to `db.internal` in the `ledger` namespace. Which outcome best explains why the secure profile limits the attack compared with the standard profile?
MCQ-13An incident review finds that a compromised production workload reached the cluster management API using a service token from its own environment. Which change isolates the management plane from the workload plane?
MCQ-6A payments service runs behind an API gateway that validates the caller's OAuth token, forwards it unchanged, strips any client-supplied X-Tenant-Id, and sets X-Tenant-Id from the token. The service trusts that header and returns the invoice named in the path. A pentest from a pod in the analytics namespace sent `curl -H 'X-Tenant-Id: 77' http://payments.svc.cluster.local:8080/v1/invoices/8842` with a token for tenant 12 and received tenant 77's invoice with HTTP 200. The nightly reconciliation job in that namespace calls payments.svc directly and must keep working. Which change closes the bypass?
MCQ-19Acme runs a multi-tenant API behind a gateway that terminates mTLS, caps each token at 100 rps, and admits inbound traffic only from the corporate VPN range. Last week a user on that VPN, holding a valid token whose tenant claim is A, sent a bodyless DELETE /api/v1/projects/9f2c. Project 9f2c belongs to tenant B. The gateway logged 200 and forwarded the call; the project service deleted the project and its 40 GB of attachments. Which control would have blocked this request?
MCQ-11A bank's transfer path has three controls: an edge gateway that validates a session JWT, a service mesh that enforces mTLS between services, and a ledger service that checks the JWT's subject against the transfer's source account. All three resolve the caller's identity from tokens signed by one auth service, whose signing key sits in a single Vault path. The gateway caches validation results for 60 seconds, and JWTs expire after 24 hours. A red-team member with the platform team's Vault token minted a fresh JWT for a customer's account and moved $2M through all three controls; the fraud model flagged the transfer 6 hours later. Which statement explains why all three controls accepted the forged token?
MCQ-12A payments API sits behind an authz sidecar. The sidecar runs a bounded fail-open: when the policy decision point (PDP) is unreachable, it serves a request only if it holds a cached decision for that session written less than 30 s ago; otherwise it denies. The PDP cluster goes down at 09:15:00. At 09:15:22 a request arrives for a session whose last cached decision, an allow, was written at 09:14:00. What does the sidecar do with this request?
MCQ-5A payments API exposes 60 HTTP routes through a public gateway. Access logs for the last 90 days show 14 of those routes received zero requests; one of the 14 proxies to the legacy billing service using a service account whose only scope is write on the ledger. The gateway's WAF carries 9 hand-written rules covering those 14 routes, and last quarter one of those rules was misconfigured for six days before anyone noticed. The team can make one change this sprint. Which change should it make?
MCQ-20An incident review of a payments platform found its break-glass admin account: excluded from the conditional-access policy so it still works when the IdP is down: had been used 4 times in 6 months. Its password, unchanged for 14 months, sits in a password manager shared by 30 engineers. Which change best closes the gap without breaking sign-in during an IdP outage?
MCQ-18A payments team cannot meet the 'hardware MFA on all admin access' requirement for 12 of its 400 admin accounts, which sit on a legacy console. It files a risk exception with a compensating control: those 12 accounts are moved behind a jump host with IP allowlisting, session recording, and 4-hour credential rotation. The vendor's MFA-capable firmware lands in Q3, five months out. The risk register scores the gap at likelihood 5, impact 4, and the control is assessed to cut likelihood to 3. The org's appetite for this category is a residual score of 10 or less. What should the risk committee do?
MCQ-23Your order platform calls a logistics vendor through one long-lived API key that grants read/write on every order and customer record. The key sits in a shared config service that all 40 backend services can read, 12 of them internet-facing. The vendor holds a current SOC 2 Type II report and a 99.9% uptime SLA. If the vendor's systems are breached next quarter, which change most reduces what the attacker can reach through your integration?
MCQ-8A team keeps nightly Postgres backups in a dedicated AWS account that production credentials cannot reach, with S3 Object Lock in compliance mode and a 14-day retention. An attacker using a stolen production admin credential dwelled 21 days before encrypting the cluster. The backup account was never touched, but its lifecycle rule had already expired every object older than 14 days, so the surviving locked copies all postdate the attacker's first write. Which single change would have left a clean, restorable copy?