CareerExamPro

CompTIA SecAI+ (CY0-001), ISACA AAISM/AAIA & EC-Council AI Security practice questions

20 questions from the SecAI+, SecOT+, AAISM/AAIA & CMMC practice test.

Start practising

CompTIA SecAI+ (CY0-001), ISACA AAISM/AAIA & EC-Council AI Security practice drill

10 realistic exam questions, graded as you go, with the exact rule or statute citation for every answer. Start free, no sign-up.

Sample questions

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?

  1. Model supply-chain compromise via serialized weights; strengthen defense by migrating model checkpoints from pickle files to cryptographically signed SafeTensors.
  2. Data poisoning via label flipping in the SIEM lake; strengthen defense by freezing training datasets and requiring two-person approval for label edits.
  3. Model extraction via surrogate querying; strengthen defense by rounding confidence probabilities to coarse tiers and throttling repeated API requests per client.
  4. 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.

Reference: CompTIA SecAI+ (CY0-001) / MITRE ATLAS (ML Attack Staging & Evasion)

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?

  1. Indirect prompt injection through transaction memos; mitigate by stripping special characters from string fields and adding a static system prompt wrapper.
  2. Vector database embedding inversion; mitigate by rotating index encryption keys and migrating embeddings from cosine similarity to Euclidean distance.
  3. Model extraction to build a surrogate replica; return coarse labels or rounded scores, apply rate limits, and flag query anomalies.
  4. 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.

Reference: CompTIA SecAI+ (CY0-001) / MITRE ATLAS (Exfiltration & Model Stealing)

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?

  1. Arbitrary code execution via pickle deserialization; mandate SafeTensors, verify signatures, scan artifacts, and load in least-privilege sandboxes.
  2. GPU side-channel memory leakage across virtual machines; mandate single-tenant bare-metal instances and disable CUDA shared-memory caching inside all build runners.
  3. Gradual weight quantization loss across container rebuilds; mandate FP32 precision exports and archive intermediate optimizer checkpoints in immutable object storage.
  4. 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)

Other topics