AI Attack Taxonomies and Classification Models COASP Practice Questions (Page 8)

Questions 36–39

1. 36application · medium

   A red team is documenting an attack against a fraud-detection model. The team first queried the model with thousands of synthetic transactions to map its decision boundary, then used that information to craft transactions that evade detection. Which taxonomy categories best describe the two phases of this attack?

   • AExtraction followed by evasion
   • BInference followed by poisoning
   • CEvasion followed by extraction
   • DPoisoning followed by inference
   Reveal Solution Discussion Select an answer first
2. 37foundation · easy

   When applying an AI attack taxonomy to a new target, what is the first step an offensive AI specialist should take?

   • AIdentify the AI components present and their attack surfaces
   • BImmediately launch a brute-force attack on the model
   • CReview the taxonomy to find the most severe attack and execute it
   • DAsk the model owner for permission to test
   Reveal Solution Discussion Select an answer first
4. 38expert · hard

   A security team is classifying vulnerabilities for a large language model (LLM) used in customer support. The team identifies two issues: (1) the model can be prompted to reveal sensitive information from its training data, and (2) the model can be tricked into generating harmful content. The team must decide which vulnerability category to report first, given that the model is deployed and the training data is proprietary. Which categorization and prioritization is most appropriate?

   • AReport both as inference-time vulnerabilities in the model component, prioritizing the data-revelation issue because it exposes proprietary information
   • BReport the data-revelation issue as a data-component vulnerability and the harmful-content issue as a model-component vulnerability
   • CPrioritize the harmful-content issue because it is more likely to be exploited than data revelation
   • DReport both as training-time vulnerabilities because the issues stem from the training data
   Reveal Solution Discussion Select an answer first
5. 39foundation · easy

   Which AI component is primarily targeted by a data poisoning attack?

   • ATraining data
   • BDeployment environment
   • CInference pipeline
   • DModel architecture
   Reveal Solution Discussion Select an answer first