10 realistic exam questions, graded as you go, with the exact rule or statute citation for every answer. Start free, no sign-up.
Timed mock exam
40 questions in 48 minutes, drawn from every topic. No answers shown until the end, then a full score breakdown, your weakest topics and a review of what you missed.
Prices are in US dollars, plus tax where it applies. Stripe, our payment processor, calculates tax at checkout.
All sales are final. If something is really wrong, such as a duplicate charge or a purchase that does not unlock what you bought, write to support@careerexampro.com and we will put it right.
One-time payment. Nothing renews.
You have an exam ticket for this practice test. Start the timed mock exam to use it.
A SOC deploys an ML malware classifier trained on PE header attributes and Windows API import tables. Without modifying the training pipeline or model weights, an adversary iteratively appends benign overlay bytes and unused legitimate DLL imports to ransomware samples until the classifier scores them as benign. Under MITRE ATLAS, how is this attack classified, and what strengthens defense?
Model supply-chain compromise via serialized weights; strengthen defense by migrating model checkpoints from pickle files to cryptographically signed SafeTensors.
Data poisoning via label flipping in the SIEM lake; strengthen defense by freezing training datasets and requiring two-person approval for label edits.
Model extraction via surrogate querying; strengthen defense by rounding confidence probabilities to coarse tiers and throttling repeated API requests per client.
Evasion via adversarial perturbation at inference time; strengthen defense with adversarial training, ensemble models, and behavioral runtime telemetry.
Show the answer
D. Evasion via adversarial perturbation at inference time; strengthen defense with adversarial training, ensemble models, and behavioral runtime telemetry.
Manipulating input features at inference time to force misclassification without altering training data or model weights is an evasion attack (MITRE ATLAS). Combining adversarial training on perturbed binaries with runtime behavioral EDR detections reduces reliance on brittle static PE features.
A corporate ERP application in the Enterprise IT network (Purdue Level 4) requires near-real-time batch yield data from a manufacturing plant Historian (Level 3), while plant PLCs (Level 1) must stay isolated from corporate hosts. How does the Purdue Enterprise Reference Architecture secure this IT/OT boundary?
Assign the ERP servers a secondary VLAN interface directly inside Level 1 so they can poll PLC registers without burdening the Level 3 Historian.
Terminate IT and OT sessions at a Level 3.5 Industrial DMZ mirror so no direct routed path connects Level 4 to Level 3.
Bridge Level 4 and Level 2 with a transparent Layer-2 switch and rely on corporate Windows Defender firewall rules to block unauthorized ERP ports.
Migrate the Level 1 PLCs into a public cloud VPC at Level 5 so corporate ERP dashboards can query field actuators over standard HTTPS APIs.
Show the answer
B. Terminate IT and OT sessions at a Level 3.5 Industrial DMZ mirror so no direct routed path connects Level 4 to Level 3.
In the Purdue Model, Level 3.5 (the Industrial DMZ) acts as a mandatory buffer between Enterprise IT (Levels 4–5) and Manufacturing/Control Operations (Levels 0–3). Replicating Historian data into a Level 3.5 mirror ensures no direct network session spans from corporate IT into plant control zones.
API telemetry shows an external B2B tenant sending 1.8 million systematically spaced synthetic feature vectors to a proprietary fraud-scoring endpoint over 48 hours, recording full floating-point class probabilities for every response. Which AI threat is most likely in progress, and which API control reduces risk without breaking legitimate customer scoring?
Indirect prompt injection through transaction memos; mitigate by stripping special characters from string fields and adding a static system prompt wrapper.
Vector database embedding inversion; mitigate by rotating index encryption keys and migrating embeddings from cosine similarity to Euclidean distance.
Model extraction to build a surrogate replica; return coarse labels or rounded scores, apply rate limits, and flag query anomalies.
Concept drift caused by seasonal spending shifts; mitigate by scheduling automated nightly model retraining on the incoming synthetic API payloads.
Show the answer
C. Model extraction to build a surrogate replica; return coarse labels or rounded scores, apply rate limits, and flag query anomalies.
High-volume, systematically distributed probing combined with collecting exact floating-point confidence vectors enables model extraction (stealing a functionally equivalent surrogate model). Truncating precision to coarse bands or binary decisions, throttling tenants, and detecting abnormal query distributions significantly slow extraction.
A power generation station asserts that its SCADA network is 'air-gapped,' yet engineers routinely transfer PLC logic via uncontrolled USB drives and a dual-homed historian exports telemetry over corporate Ethernet. Which control physically enforces outbound-only telemetry export while preventing inbound cyber commands?
A hardware unidirectional security gateway (optical data diode) with no return signal path, paired with a hardened removable-media scanning kiosk.
A standard stateful next-generation firewall configured with an outbound-only ACL rule while keeping full bidirectional copper Ethernet cabling attached.
A software VPN concentrator running TLS 1.3 between the dual-homed historian and corporate analytics servers over an untagged trunk port.
A corporate NAC policy that assigns contractor USB flash drives a temporary guest IP address whenever plugged into engineering workstations.
Show the answer
A. A hardware unidirectional security gateway (optical data diode) with no return signal path, paired with a hardened removable-media scanning kiosk.
True air gaps rarely exist once transient USB media and dual-homed hosts appear. A hardware data diode uses a transmit-only LED and receive-only photodiode to physically eliminate any inbound return channel (unlike software firewalls that can be misconfigured or exploited), while a kiosk sanitizes removable media.
An MLOps CI/CD runner pulls third-party open-weights checkpoints in legacy PyTorch .pkl/.pt format from a community model hub and deserializes them directly inside a GPU container attached to a broad production IAM role. What is the primary supply-chain vulnerability, and what MLSecOps standard prevents exploitation?
Arbitrary code execution via pickle deserialization; mandate SafeTensors, verify signatures, scan artifacts, and load in least-privilege sandboxes.
GPU side-channel memory leakage across virtual machines; mandate single-tenant bare-metal instances and disable CUDA shared-memory caching inside all build runners.
Gradual weight quantization loss across container rebuilds; mandate FP32 precision exports and archive intermediate optimizer checkpoints in immutable object storage.
Unintended hallucination of deprecated library APIs; mandate deterministic decoding with temperature zero and enable automated unit-test retries in CI/CD.
Show the answer
A. Arbitrary code execution via pickle deserialization; mandate SafeTensors, verify signatures, scan artifacts, and load in least-privilege sandboxes.
Python pickle files execute embedded deserialization routines (such as __reduce__) upon loading, allowing malicious model weights to run arbitrary shell commands with the container's IAM privileges. Enforcing non-executable formats like SafeTensors, verifying cryptographic hashes/signatures, and isolating model loaders with least privilege eliminates this attack path.
Reference: CompTIA SecAI+ (CY0-001) / OWASP Top 10 for LLM Applications (LLM03: Supply Chain)