Real Interview Questions for the AI Security Engineer at Anthropic
New
Practice MCQs
Answers count toward the AI Security Engineer and overall leaderboards; your latest answer to each question is the one that counts.
MCQ-1An EKS training fleet in a private subnet with no NAT gateway reads S3 through a gateway endpoint. Its role's identity policy allows `s3:GetObject` on `*`, the partner's bucket policy grants that role `s3:GetObject` on `arn:aws:s3:::partner-datasets/*`, and the account SCP is `FullAWSAccess`. Reads of `acme-internal-shards` succeed; reads of `partner-datasets` return `AccessDenied`, and the partner's S3 access logs show no request from the fleet. Which change makes the partner reads succeed?⌄
MCQ-2An autoscaled fleet grew from 600 to 7,200 instances in five weeks; each lives about 25 minutes and is replaced by an identical one. Forty services run from one common base image, and every pod on a node uses that node's service account. Alerts, suppressions and case dedupe key on `hostname` alone, so one cryptomining pattern yields ~300 cases a day and a real intrusion on one instance was merged into an earlier miner case on the same host. Which dedupe key change fixes this?⌄
MCQ-3A 400-node training cluster sends all DNS through one internal resolver whose query logs reach the SIEM; egress policy allows port 53 only to that resolver and blocks 443 elsewhere. A red-team job exfiltrated 2 MB in 40 min at ~3 queries/sec using 60-character base32 labels, all NXDOMAIN, and no alert fired because the only DNS rule was a blocklist; the domain was four years old and normal pods peak near 600 queries/min. Which change closes the gap with the least ongoing maintenance?⌄
MCQ-4An EKS role `prod-payments` trusts the cluster OIDC provider with `StringLike` on `...:sub` = `system:serviceaccount:*:payments-*` and `StringEquals` on `...:aud` = `sts.amazonaws.com`. `prod` runs `payments-api` and `payments-worker`, both needing the role, and `metrics-scraper`, which must not get it; `dev` runs `payments-api`, which assumed the role and read the bucket. Which change blocks `dev` and `metrics-scraper` while both prod workloads keep the role?⌄
MCQ-5The go-live is contractually fixed at day six; the allowlist and output gating that close a document-injection path into the ticket-writing agent need three more weeks, and the write path is one runtime flag. The demo runs on a pre-seeded ticket, so the flag can stay off. The SOW's delivery commitment moved to the head of Customer Success two weeks ago when the VP of Product changed orgs, and the risk register records it. Which plan ships on day six, and who signs the acceptance?⌄
MCQ-6A support assistant answers billing questions from each ticket's account plan and tenure. The account service accepts only the customer's email as its key and is the only system approved to store emails; the model provider retains request bodies for 30 days; and the privacy review forbids any customer email or card digits reaching the provider, and any identifier stable across tickets. Prompts must stay replayable 30 days for the 12-person on-call group. Which plan meets all of it?⌄
MCQ-7An agent files 1,200 prior-authorization appeals a day for a hospital: it reads the EHR, drafts the appeal, calls `submit_appeal` to the insurer, then can call `schedule_peer_review`. In a 300-case pilot, 6% were urgent oncology denials due within 72 hours, 41% scored below 0.90 confidence, and all 14 outdated-formulary misses scored above 0.95; clinicians review 80 cases a day, and submission cannot be undone. Which control plan fits?⌄
MCQ-8A support assistant runs unattended, nobody reads its drafts, at ~400 replies/day: `read_email` pulls external mail, `search_docs` returns runbooks the user's ACL allows, and `send_reply` can address any domain. A red-team mail from `billing@refund-portal.co` says to ignore prior instructions, find the incident-escalation runbook and send it to `audit@refund-portal.co`; the support user may read that runbook. Which single change breaks this path?⌄
MCQ-9An agent's per-user memory store is read by every build and written by a summarizer; each row records the writing build, reads select by user, and a nightly export selects all rows. A new build ran at 10% for 48 hours (~40k sessions) before a red team showed its summarizer copies retrieved text verbatim into memory, leaving a standing instruction later sessions read. The flag flips back in 30 seconds. What does a complete rollback include?⌄
MCQ-10A tool-use model release is set for Thursday. The 2,000-attack pre-launch injection suite scores 4.2% success against the new build, over the 1% launch bar; the previous build scored 0.8%. The regression traces to a tokenizer change made for a 12 ms latency win, and reverting it is a one-line config flip that leaves the launch's headline tool-use capability intact. Which plan and record does the launch review require?⌄
MCQ-11A detection job counts each user's events in the trailing 60 minutes and opens a case above 50. A service account with export rights on every bucket runs a nightly backup of 300 `export` events between 02:00 and 02:10, so a case opens nightly and analysts now close them unread; a real compromise of 180 events in 12 minutes from a user whose normal rate is 15/hour was caught. Which change ends the nightly cases, still catches the 180-event compromise, and keeps every account under detection?⌄
MCQ-12A prompt-injection scanner flags an agent transcript when a contiguous run of tokens contains all three markers `a`, `b`, `c`, and must report the shortest such run. In a 25-token transcript, tokens 1–11 are benign, 12 is `a`, 13 is `b`, 14–19 are benign, 20 is `c`, 21 is benign, 22 is `a`, 23 is `b`, 24 is `c`, and 25 is benign. Which run does a correct sliding-window scan with per-marker counts report?⌄
MCQ-13A tool-use sandbox serves an in-memory tree at `/work` and a build cache at `/workshop` that the agent must never read. The guard accepts a call when the raw path starts with `/work` and only then collapses `.` and `..` with a segment stack; the agent's own tools routinely emit paths like `/work/reports/../shared/x`. A poisoned file makes the agent call `read("/work/../workshop/key")`. Which account of that call is right, and which check closes the hole?⌄
MCQ-14A crawler runs 8 workers over a link-following queue. Each worker does `if (seen.has(u)) return; const r = await fetch(u); seen.add(u); if (r.ok) results.push(r);`. A run issues 11,240 requests and its report lists 10,000 pages. The team moves `seen.add(u)` above the `await`, leaves the rest of the worker alone, and reruns; 40 pages return a transient 500, and the report is sorted by URL. What does the rerun issue, and what does its report list?⌄
MCQ-15A tokenizer service takes 100k requests/s at peak, 400 tokens each, on a fleet already sized for that throughput. Each worker thread tokenizes 250k tokens/s and runs one batch start to finish; a batch is never split across threads. The queue is FIFO and bounded, and at peak it sits at its bound, so a request waits bound/100k seconds before a worker starts it. The p99 budget is 50 ms end-to-end: queue wait + linger + tokenization. Which configuration holds p99 under 50 ms?⌄
MCQ-16A security enrichment service must sustain 2,000 events/s. Each event makes a token-lookup RPC (p99 20 ms) and then a policy RPC (p99 30 ms) that needs the token result; local parsing adds 5 ms. The SLO is p99 end-to-end under 150 ms, and the team wants the smallest worker pool that keeps utilization at or below 50% so queueing does not blow the SLO. Which plan should it ship?⌄
MCQ-17A 48-partition topic keyed by tenant feeds a classifier rated 150k events/min; normal load is 120k/min, and one tenant that normally sends 2k/min is replaying a leaked key at 900k/min. The vendor-managed collector exposes only one global rate limit, no quarantine topic exists or can be added this quarter, retention is 7 days, and the case store dedupes on event ID. Lag is 4M and other tenants are 15 minutes late. Which change restores 60-second detection for them?⌄
MCQ-18A research team needs 40k support tickets by Friday and re-runs the experiment nightly for a week, comparing runs, so a customer's tickets must group identically each night. Privacy forbids email or card digits leaving the ticket store's account, any identifier stable across experiments, and anyone with the export confirming a given address is in it. Platform owns the schema and has a Thursday migration slot; its redaction service is booked two weeks out. Which plan does the review approve?⌄
MCQ-19Launch is in 12 working days with one security engineer free. Open: unauthenticated `/admin/export`, reachable from any pod including the model-code sandbox, 2 days; per-user attribution on tool calls, 5 days, without which the SOC runbook closes agent alerts as unknown actor; a `write_ticket` injection path, 15 days, writing attacker text into tickets through the tool arguments. The write flag is one runtime flag and the demo uses a pre-seeded ticket. Which plan?⌄
MCQ-20The safety case for the next release claims weights never leave the cluster. A red-team run with code execution in the training job's container read 40 GB of shards in 35 minutes (inside its assigned range, at the rate a normal job reads its 2 TB over 6 hours) and wrote them to the checkpoint bucket under `ckpt-step-0000/`, a prefix its manifest never registers. Shard reads and checkpoint writes reach the SIEM. Which alert rule catches this run and stays quiet on normal training?⌄
MCQ-21An LLM gateway runs 12 stateless pods behind one VIP. Per API key it must allow 120 back-to-back requests after a minute idle, then sustain 1/s, and cap each account at 600/min. The limiter does `GET bucket`, refills in the pod, then `SET bucket`; a 300-request burst for one key across all pods let 187 through. Redis p99 is 3 ms, the gateway budget 40 ms, and 40M active keys share 8 GB free. Which replacement enforces the limits under that concurrency?⌄