AI Vulnerability Scanning and Fuzzing COASP Practice Questions (Page 3)

Questions 11–15

1. 11application · medium

   A security team is fuzzing a speech recognition model. They want to use a mutation-based strategy that is tailored to the audio input space. Which approach best fits this requirement?

   • ARun the model with the same audio sample repeatedly to check for consistency.
   • BMutate the raw audio waveform by adding small amounts of noise and observe changes in transcription.
   • CRandomly replace words in the transcript with synonyms and test the model's response.
   • DUse a coverage-guided fuzzer that only measures code coverage of the underlying C++ library.
   Reveal Solution Discussion Select an answer first
2. 12application · medium

   A company is deploying a new AI system that uses a third-party pre-trained model, a custom data preprocessing pipeline, and a cloud-hosted inference API. The security team must map the attack surface before scanning. Which component is most likely to be overlooked but is critical for AI-specific vulnerabilities?

   • AThe SSL/TLS certificate of the API endpoint
   • BThe data preprocessing pipeline that transforms raw input before it reaches the model
   • CThe cloud provider's physical data center
   • DThe web application firewall protecting the API
   Reveal Solution Discussion Select an answer first
4. 13expert · hard

   A security team is fuzzing a malware detection model that takes executable files as input. The team needs to generate diverse test cases that are realistic enough to bypass the model's detection but also cover a wide range of malware families. They have a limited compute budget and need to prioritize the most promising inputs. Which approach best balances diversity, realism, and efficiency?

   • AStart with a corpus of known malware and apply mutation-based fuzzing to create variants, using a coverage-guided strategy to prioritize inputs that exercise new model paths.
   • BRandomly generate byte sequences and feed them to the model without any filtering.
   • COnly use the original malware corpus without any mutations to avoid introducing unrealistic samples.
   • DUse a generative adversarial network (GAN) to create new malware samples from scratch, then test them all.
   Reveal Solution Discussion Select an answer first
5. 14foundation · easy

   What is the primary goal of coverage-guided fuzzing in AI?

   • ATo reduce the model's training time
   • BTo ensure the model's outputs are always correct
   • CTo minimize the number of test cases
   • DTo maximize the exploration of the model's internal states or decision boundaries
   Reveal Solution Discussion Select an answer first
6. 15expert · hard

   A fuzzing campaign on a recommendation system produces a large number of inputs that cause the model to recommend inappropriate content (e.g., violent videos to minors). The team has limited resources and must decide which findings to fix first. Which prioritization approach is most appropriate?

   • AIgnore the findings because fuzzing inputs are not representative of real user behavior.
   • BPrioritize findings that are likely to occur in real user behavior and have the highest safety impact, such as those involving common input patterns.
   • CFix all findings equally, since any inappropriate recommendation is unacceptable.
   • DFocus only on findings that cause system crashes or errors, not content issues.
   Reveal Solution Discussion Select an answer first