AI Attack Surfaces, Threat Landscapes, and Adversary Techniques (MITRE ATLAS-aligned) COASP Practice Questions (Page 6)

Questions 26–30

1. 26application · medium

   A red team is documenting an attack where they used a publicly available model card to identify the model architecture, then crafted inputs that were misclassified. Which MITRE ATLAS tactic and technique pair best describes the use of the model card?

   • AReconnaissance via Search Open Technical Databases
   • BEvasion via Adversarial Inputs
   • CInitial Access via Valid Accounts
   • DExfiltration via ML Model Inference
   Reveal Solution Discussion Select an answer first
2. 27foundation · easy

   Which MITRE ATLAS technique is most directly associated with the 'data' attack surface of an AI system?

   • AModel Inversion
   • BAdversarial Examples
   • CData Poisoning
   • DDenial of Service
   Reveal Solution Discussion Select an answer first
4. 28expert · hard

   A security engineer is using MITRE ATLAS to assess the risk of an ML system that uses a third-party API for sentiment analysis. The engineer is concerned about the possibility of an attacker using the API to extract the model's training data. Which ATLAS technique is most relevant to this concern?

   • AExtract ML Model
   • BInfer Training Data Membership
   • CPoison Training Data
   • DExfiltration via Cyber Means
   Reveal Solution Discussion Select an answer first
5. 29foundation · easy

   Which MITRE ATLAS technique is most directly associated with the 'deployment' attack surface of an AI system?

   • ADenial of Service
   • BModel Inversion
   • CData Poisoning
   • DAdversarial Examples
   Reveal Solution Discussion Select an answer first
6. 30application · medium

   A security analyst is documenting a real-world attack where an attacker added imperceptible noise to a stop sign image, causing an autonomous vehicle's object-detection model to classify it as a speed limit sign. The analyst wants to map this to MITRE ATLAS. Which tactic and technique pair is most appropriate?

   • AML Model Inference – Extract ML Model
   • BML Attack Staging – Craft Adversarial Data
   • CInitial Access – Exploit Public-Facing Application
   • DML Model Inference – Adversarial Input
   Reveal Solution Discussion Select an answer first