AI Attack Taxonomies and Classification Models COASP Practice Questions (Page 6)

Questions 26–30

1. 26foundation · easy

   An AI attack taxonomy is best described as which of the following?

   • AA set of defensive tools used to detect and mitigate attacks against AI systems
   • BA structured framework that categorizes attacks based on their target, attacker goal, and stage in the AI lifecycle
   • CA list of all known vulnerabilities in a specific AI model, ranked by severity score
   • DA legal framework that defines penalties for unauthorized access to AI systems
   Reveal Solution Discussion Select an answer first
2. 27application · medium

   A security team is classifying vulnerabilities found in their AI-powered recommendation engine. They have identified two issues: (1) an attacker can query the model to determine if a specific user's purchase history was used in training, and (2) an attacker can submit crafted inputs that cause the model to recommend inappropriate products. Which vulnerability categories and attack types do these issues represent, respectively?

   • AInference vulnerability (membership inference) and evasion vulnerability (adversarial examples)
   • BData vulnerability (poisoning) and model vulnerability (extraction)
   • CInference vulnerability (model inversion) and data vulnerability (poisoning)
   • DModel vulnerability (extraction) and inference vulnerability (membership inference)
   Reveal Solution Discussion Select an answer first
4. 28foundation · easy

   How does an offensive AI practitioner best use an attack taxonomy to prioritize techniques?

   • ABy avoiding any technique not explicitly listed in the taxonomy
   • BBy using the taxonomy only to document attacks after they have occurred
   • CBy mapping each technique to the target AI component and attacker goal, then ranking by likelihood and impact
   • DBy selecting the first technique listed in the taxonomy for every engagement
   Reveal Solution Discussion Select an answer first
5. 29application · medium

   A security analyst is investigating a breach of a recommendation system. The analyst finds that the attacker had access to the model's training dataset and inserted fake user profiles that skewed the model's recommendations toward the attacker's products. Which attack category and lifecycle stage does this represent?

   • APoisoning attack during the training stage
   • BInference attack during the post-deployment stage
   • CExtraction attack during the deployment stage
   • DEvasion attack during the inference stage
   Reveal Solution Discussion Select an answer first
6. 30expert · hard

   An offensive AI team is planning a test for a chatbot model. The team has two objectives: (1) test the model's resistance to prompt injection attacks, and (2) test the model's resistance to data-extraction attacks. The team has limited time and must choose which attacks to test first. The model is deployed and the training data is highly sensitive. Which prioritization is most defensible?

   • ASkip both attack types because the model is deployed and cannot be changed
   • BTest both attack types simultaneously because they are equally important
   • CPrioritize prompt injection attacks because they are easier to execute than extraction attacks
   • DPrioritize data-extraction attacks because the training data is highly sensitive and exposure would have severe consequences
   Reveal Solution Discussion Select an answer first