AI Attack Taxonomies and Classification Models COASP Practice Questions (Page 5)

Questions 21–25

1. 21application · medium

   A security team is classifying a vulnerability report for a natural-language-processing model. The report states that an attacker can submit a specially crafted sentence that causes the model to output a toxic response, even though the model was fine-tuned with safety filters. The model's weights and training data are unchanged. Which vulnerability category and attack surface does this report describe?

   • AExtraction vulnerability in the inference component
   • BInference-time vulnerability in the model component
   • CTraining-time vulnerability in the data component
   • DDeployment vulnerability in the infrastructure component
   Reveal Solution Discussion Select an answer first
2. 22application · medium

   A machine learning operations (MLOps) team is conducting a security review of their model deployment pipeline. They are categorizing risks based on where in the AI system a vulnerability could be exploited. Which of the following correctly categorizes a vulnerability in the model's decision-making logic that can be triggered by specific input patterns?

   • AModel vulnerability; exploitation via evasion attack
   • BInference vulnerability; exploitation via extraction attack
   • CData vulnerability; exploitation via poisoning attack
   • DInfrastructure vulnerability; exploitation via denial of service
   Reveal Solution Discussion Select an answer first
4. 23application · medium

   An offensive AI team is planning a penetration test against a healthcare organization's diagnostic model. The team wants to systematically identify which attacks are most relevant given the model's high regulatory impact and the attacker's limited access. Using an AI attack taxonomy, which approach best supports this prioritization?

   • AList all known attack types from the taxonomy and test them in alphabetical order
   • BFocus exclusively on poisoning attacks because they are the most severe category in all taxonomies
   • CSelect only inference attacks because they are the only ones that apply to deployed models
   • DMap each potential attack to the affected AI component and the attacker's required access level, then rank by regulatory impact
   Reveal Solution Discussion Select an answer first
5. 24expert · hard

   A penetration tester is using an AI attack taxonomy to guide an assessment of a healthcare AI system that predicts patient readmission. The tester has a limited budget and must choose between testing for (a) model extraction to steal the proprietary model, or (b) evasion attacks to cause incorrect predictions. The system is deployed behind an API with rate limiting, and the tester has no insider access. Which approach is more aligned with a systematic taxonomy-driven assessment, considering the constraints?

   • APrioritize data poisoning, as it is the most severe attack type in the taxonomy
   • BPrioritize both equally, as the taxonomy requires a comprehensive assessment of all attack types
   • CPrioritize evasion attacks, as they can be developed and tested against the API with a limited number of queries, and the impact is directly aligned with the system's purpose
   • DPrioritize model extraction, as it provides the most comprehensive understanding of the system for future attacks
   Reveal Solution Discussion Select an answer first
6. 25foundation · easy

   Which of the following is a primary purpose of using an AI attack taxonomy in offensive security?

   • ATo automatically patch vulnerabilities in AI models without human intervention
   • BTo provide a common language for describing and comparing AI attacks across different systems
   • CTo replace the need for penetration testing by providing a complete list of all possible attacks
   • DTo guarantee that all AI systems are secure if they follow the taxonomy's recommendations
   Reveal Solution Discussion Select an answer first