AI Vulnerability Scanning and Fuzzing COASP Practice Questions (Page 5)

Questions 21–25

1. 21application · medium

   A security analyst is using a coverage-guided fuzzer to test a deep learning model for image classification. The fuzzer reports that it has achieved high neuron coverage, but the model still misclassifies many generated inputs. What is the most likely reason for this discrepancy?

   • ANeuron coverage is not a perfect proxy for model robustness; the fuzzer may be exploring areas of the input space that do not correspond to realistic adversarial perturbations.
   • BThe fuzzer is not using enough CPU cores to generate diverse inputs.
   • CThe fuzzer is incorrectly measuring neuron coverage due to a bug in the instrumentation.
   • DThe model is too small to be adequately tested by coverage-guided fuzzing.
   Reveal Solution Discussion Select an answer first
2. 22application · medium

   During a vulnerability scan of a recommendation system, the security team discovers that an attacker can query the model API with carefully crafted inputs and infer whether a specific user's data was used in training. Which AI-specific vulnerability category does this finding represent, and what is the most appropriate initial mitigation?

   • AData poisoning; retrain the model with a clean dataset.
   • BAdversarial perturbation; apply input sanitization to block malicious queries.
   • CMembership inference; implement differential privacy during training and add rate limiting to the API.
   • DModel inversion; restrict API access to authenticated users only.
   Reveal Solution Discussion Select an answer first
4. 23application · medium

   A security engineer is tasked with performing an initial vulnerability scan of a newly deployed AI-powered fraud detection system. The system uses a gradient-boosted model trained on historical transaction data, with a REST API for real-time scoring. The engineer wants to identify weaknesses that are unique to the AI components, not just the underlying web server. Which approach best addresses this requirement?

   • APerform a model-level assessment that includes testing for adversarial perturbations, membership inference, and data poisoning indicators, in addition to scanning the API and data pipeline.
   • BConduct a network vulnerability scan of the hosting infrastructure and verify TLS configurations.
   • CReview the model's training code for hardcoded credentials and dependency vulnerabilities.
   • DRun a standard web application scanner against the REST API and patch any OWASP Top 10 issues found.
   Reveal Solution Discussion Select an answer first
5. 24foundation · easy

   Which AI vulnerability category involves an attacker inserting malicious samples into the training dataset to corrupt the model's behavior?

   • AMembership inference
   • BAdversarial perturbations
   • CData poisoning
   • DModel inversion
   Reveal Solution Discussion Select an answer first
6. 25application · medium

   A security team is fuzzing a natural language processing (NLP) model that classifies customer reviews as positive or negative. They want to maximize the discovery of edge-case failures while keeping the fuzzing campaign efficient. Which strategy best balances coverage and efficiency for this NLP model?

   • ARun a fixed set of handcrafted adversarial examples and stop after they are tested.
   • BUse a coverage-guided fuzzer that measures neuron activation and prioritizes inputs that exercise new paths.
   • COnly test the model with the most common review phrases to ensure the main paths work.
   • DRandomly mutate characters in the reviews without tracking which parts of the model are exercised.
   Reveal Solution Discussion Select an answer first