Skip to content
Tech Interview Prep
Technical interview guide

Threat Modeling & Risk Assessment

Systematically identifying what could go wrong, before it does: frameworks like STRIDE and risk scoring.

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

Scope: NIST SP 800-30 Rev. 1; NIST CSF 2.0; NIST SP 1303; NIST SP 800-154 initial public draft status noted; ATT&CK and CAPEC accessed 2026-08-31.

Curated: · Written: · Reviewed:

25 questions

MCQ-6A team is filling a risk register with two columns: vulnerabilities and threat events. Which entry belongs in the vulnerability column?
MCQ-12A payments team builds an attack tree for the goal "transfer funds without authorization." The root is an OR node with three children: (1) an AND node "compromise an operator workstation AND reuse its cached session token," (2) a leaf "phish an operator's credentials," and (3) a leaf "exploit an unpatched API endpoint." They ship two changes: session tokens are bound to the workstation's TPM so they cannot be replayed elsewhere, and operators move to phishing-resistant FIDO2 keys. The API endpoint stays unpatched. What does the updated tree show about the attacker's options?
MCQ-15A payments team's threat model lists only preventive controls: signed builds, mTLS between services, per-service IAM roles. Their last two incidents were reported by customers, and in both the attacker had been inside for over a week. The team adds detection and recovery entries to the model. What do those entries contribute that the preventive ones do not?
MCQ-1A team runs a two-hour STRIDE session on a new refund service and gets 38 one-line candidate threats spread across all six categories. Launch is in five days and the team can fix at most four issues. Which action best follows from what the session produced?
MCQ-3A team is threat modeling a document-upload feature before launch. The browser asks the upload service for a pre-signed S3 URL, then PUTs the document bytes straight to an S3 bucket encrypted with SSE-KMS. S3 emits an event to SQS; a worker reads the object, writes metadata to Postgres, and sends the full document to a third-party OCR API. Login is delegated to a third-party OIDC provider. The team wants the smallest scope that still covers every trust boundary the document bytes or the login decision crosses. Which components belong in the model?
MCQ-4A team reviews the data-flow diagram for its checkout API: mobile app → API gateway → order service → SQS queue → settlement worker → Postgres. The gateway terminates TLS, validates the JWT, strips it, and forwards the request to the order service with `X-User-Id` set from the token's `sub` claim; the order service reads that header and uses it as the authenticated caller for authorization. Gateway-to-service traffic runs over the mesh with mTLS. The order service's port is reachable from all 14 pods in the cluster namespace. The queue uses SSE-KMS, the worker has its own IAM role scoped to the settlement tables, and Postgres accepts connections only from the worker's security group. Which trust-boundary defect should the review record?
MCQ-5`token-svc` is an internal gRPC service that answers card-token lookups. It is reachable only from `checkout-svc` inside the VPC, authenticates callers with mTLS over TLS 1.3, reads from a read-only Postgres replica, and exposes a single `Lookup(token_id)` RPC. During threat modeling, the team wrote the entries below. Which one is a testable scenario for this system?
MCQ-2A team threat-models a webhook receiver. The DFD is the payment provider as an external entity, an HTTPS data flow into the receiver, the receiver process, and an events table it writes to. The receiver already verifies the HMAC signature on every request, and the gateway rate-limits to 100 requests per second. During review, an engineer replays a webhook captured from yesterday's traffic, byte-for-byte unchanged, and the receiver credits the account a second time. Which control should the threat model add?
MCQ-13An internal refund console lets support agents issue refunds. A refund above $500 needs approval from a supervisor other than the requester; each agent is capped at 60 refunds per hour; refunds to merchant accounts on the fraud blocklist are rejected; and a transaction can be refunded only once. Every refund is written to an audit log with agent ID, amount, timestamp and merchant account. During threat modeling the team is listing abuse cases that these controls do not address. Which of the following belongs on that list?
MCQ-22A threat model lists this control for the login service: "Lock the account for 15 minutes after 5 failed password attempts within 10 minutes." To verify it, an engineer runs a script against a staging account whose password is known: five requests with a wrong password, then a sixth request with the correct password. The five wrong-password requests each return 401; the sixth returns 200 and a session token. What does this result establish about the control?
MCQ-21A payments team is adding a fraud-scoring vendor. The service currently sends only tokenized card IDs to external services. The new integration will send raw PANs over a new outbound mTLS connection to the vendor, and the vendor will call a new webhook endpoint on the payments service using a shared secret. The existing threat model was completed 10 months ago and assumes PANs never leave the tokenization boundary and that all inbound calls come from the internal API gateway. The vendor has a current SOC 2 Type II report. What should the security engineer do before the integration ships?
MCQ-23A payments team runs an LLM threat-generation pass over the design doc of a new service and gets 400 candidate threats back. Deduplicating by hand, a reviewer finds 160 of the 400 are restatements of a candidate already on the list, leaving 240 distinct threats. She then checks each distinct threat against the service's current data-flow diagram: 145 of the 240 name at least one component, data store or trust boundary that appears nowhere in the diagram, and the names trace back to a public reference architecture the model saw in training. The design freeze is at the end of the week and the security architect has one 90-minute session available. How many candidates should go into that session?
MCQ-20A team threat-models the dose-entry service for a hospital's infusion pumps. The STRIDE pass finds every endpoint behind mTLS, no injection or privilege-escalation path, and no unencrypted PHI at rest; the model closes with 'no high-severity threats.' Two months later a nurse types 250 into a field labelled mcg that the pump firmware reads as mg (an authenticated, authorized, well-formed request) and the pump delivers a 1000x overdose. The review rules the threat model should have flagged this. Which single change to the threat model would have surfaced this finding?
MCQ-25A payments team threat-models a service and records five scenarios: | # | Scenario | Likelihood | Impact | Status | | --- | --- | --- | --- | --- | | S1 | Stolen API key reads another tenant's data | 4 | 5 | Mitigated: per-tenant scoping shipped | | S2 | Replayed signed webhook | 3 | 4 | Accepted: control depends on the vendor's timestamp check, not yet tested | | S3 | SSRF from the image-fetch endpoint | 5 | 4 | Open: new control still in review | | S4 | Insider reads the audit log | 2 | 5 | Accepted: named owner, no third-party dependency | | S5 | Credential stuffing on login | 4 | 3 | Mitigated: rate limit shipped | Policy: escalate to the review board any open scenario with likelihood ≥ 4 and impact ≥ 4, and any accepted scenario whose residual risk depends on an unverified third-party assumption. The model's conclusion must flag the scenarios that go to the board. Which scenarios does the conclusion flag?
MCQ-9A payments team quantifies credential-stuffing fraud on its login API. The inherent annualized loss expectancy is $400,000. A bot-detection control covers 75% of the exposure and telemetry shows it cuts risk by 90% on that portion; the remaining 25% of the exposure is unaffected. What residual annualized loss expectancy should the risk register record?
MCQ-11A team is sizing schedule risk for a database migration. The three-point estimate is optimistic 4 days, most likely 6 days, pessimistic 20 days; the pessimistic case assumes undocumented triggers in the legacy schema. The release window is 10 days. Which entry should the risk register carry?
MCQ-10A payments team scores each open risk on a 5-point likelihood scale and a 5-point impact scale, then ranks by the product. Two risks are open: | risk | likelihood | impact | score | | --- | --- | --- | --- | | A | 5 | 2 | 10 | | B | 3 | 4 | 12 | B is ranked first. A reviewer proposes re-anchoring the likelihood scale to 1, 2, 4, 8, 16: the same five levels in the same order, spaced differently. Under that scale A scores 16 × 2 = 32 and B scores 4 × 4 = 16, so A is ranked first. What does the reversal show?
MCQ-7A CVSS v3.1 base score of 9.8 (exploitability subscore 3.9, vector AV:N/AC:L/PR:N/UI:N) is filed against a library compiled into your service. The service listens only on an internal VLAN, and reaching it requires an MFA-authenticated jump host. The vulnerable function sits behind a feature flag that is off in every production configuration, and no public exploit exists. What likelihood belongs in this quarter's risk register for this finding?
MCQ-8A municipal water utility runs its lab information management system (LIMS) on the corporate network. A unidirectional gateway blocks all traffic from the corporate network into the OT network, so the LIMS cannot send commands to plant controllers. The utility has no inline chlorine analyzer: control-room operators read the LIMS dashboard's chlorine residual readings every two hours and set the dosing pumps from those numbers. A penetration test shows that any authenticated LIMS user can alter the stored readings and what the dashboard displays. The utility's risk rubric scores impact as the highest of three dimensions, each 1–5. Safety is 5 if falsified data can lead an operator to set a physical process parameter to an unsafe value, 3 if it can only delay or misroute a response, 1 otherwise. Regulatory is 5 if personal data on more than 1,000 people is exposed, 3 for 100–1,000, 1 otherwise. Availability is 5 if the outage exceeds 24 hours, 3 for 1–24 hours, 1 otherwise. The finding touches 40 employee records, and restoring the LIMS from backup takes about 30 minutes. What impact score should the register carry?
MCQ-17A quarterly access review finds 14 service accounts in the payments platform still holding standing admin rights. The risk register entry scores likelihood 3 × impact 4 = 12 on the organization's 5×5 matrix. The risk appetite statement defines three acceptance bands: 1–10 may be accepted by the service's named risk delegate; 11–16 by the accountable business owner for the service; 17–25 by the CIO. A delegate counts only if named in the register entry. Remediation, moving the accounts to just-in-time elevation, is funded and lands in 9 weeks. The platform's named risk delegate wants to sign the acceptance now so the finding closes this quarter. Which signature closes the finding within these rules?
MCQ-18`checkout-api` (Node 20, npm 10) depends on `@acme/logger@4.0.0`, which pulls in `fast-parse@2.3.0`. SCA flags a CVSS 9.1 prototype-pollution RCE in `fast-parse`'s `parseConfig()`. The team confirms `parseConfig()` is called only from `src/devtools/seed.ts`, which `tsconfig.build.json` excludes, so the function never runs in the deployed service. `fast-parse` also ships a `postinstall` script (`node scripts/fetch-dict.js`) that runs during `npm ci` in the CI job, and that job assumes `AWS_DEPLOY_ROLE`, which can push to the production image registry and read production secrets. Which assessment is most defensible?
MCQ-19A payments service is built in CI by this Dockerfile: FROM node:20-slim ARG DB_PASSWORD ARG FRAUD_API_KEY WORKDIR /app COPY package*.json ./ RUN npm ci COPY . . CMD ["node","server.js"] CI passes both values with `--build-arg`. The two ARGs are left over from a build step that was deleted months ago; the running container gets both secrets from its environment instead. `npm ci` installs 40 transitive packages from a lockfile in 41 seconds and runs those packages' own install scripts. The pipeline failed at the test stage, so the image was never pushed anywhere. The build host's egress log shows one connection to 203.0.113.47, an address the team does not own, during those 41 seconds, and six minutes later the DB password from that job was used to log in to Postgres from the same address. Which finding should the security review record?
MCQ-14A payments team's threat model ranks attack paths by annualized loss exposure (ALE). The top three are S1, credential stuffing leading to account takeover, $480k; S2, SSRF from the image-resizer reaching the instance metadata service, $300k; and S3, an API key leaked in a public repo, $120k. A reflected-XSS path on the marketing site ranks 14th at $15k. The team can ship exactly one control this quarter. Pilot and vendor data: - X: rate limiting plus device fingerprinting on the login endpoint. A pilot on 10% of traffic cut successful account takeovers by 70%. - Y: IMDSv2 enforcement plus an egress allowlist on the image-resizer. A staging pilot eliminated 95% of SSRF attempts that reached the metadata service. - Z: a vendor WAF rule pack marketed as covering the OWASP Top 10. No pilot; the vendor publishes no measured effect on S1–S3. - W: a reflected-XSS fix on the marketing site. A pilot eliminated 99% of that path's exploitation attempts. - V: a shared authentication library upgrade. The threat model estimates it would cut S1 by 25% and S3 by 40%; no pilot has been run. Which control should the team ship?
MCQ-16A payments team has three engineer-weeks of remediation capacity before its quarterly audit. Four findings are open, each with an estimated annualized loss expectancy (ALE) and a fix effort: | finding | ALE | fix effort | basis | | --- | --- | --- | --- | | F1: token-signing key shared across tenants | $1.2M | 3 wk | one pen-test scenario; the platform team puts the shared-key assumption at 30% likely | | F2: credential stuffing on the public login | $300k | 1 wk | 40k attempts/day in the logs, 0.3% succeed | | F3: stack traces in public API error responses | $120k | 1 wk | reproduced against production | | F4: session replay after logout | $60k | 1 wk | reproduced in staging | Which allocation of the three engineer-weeks is most defensible?
MCQ-24A security team is choosing a single metric to report to the CISO for its threat-modeling program. The metric will be used to compare teams and determine bonuses. Which metric most risks being gamed?