AI Reconnaissance and Attack Surface Mapping COASP Practice Questions (Page 3)

Questions 11–15

1. 11application · medium

   You have completed reconnaissance on an AI system and identified the following: a public API, a model trained on proprietary data, and a deployment on a cloud VM. Which of the following is the most complete attack surface map?

   • AA recommendation to use a Web Application Firewall (WAF) to protect the API.
   • BA report on the model's accuracy and performance metrics.
   • CA list of attack vectors for the API (e.g., adversarial inputs), the model (e.g., model inversion), and the VM (e.g., OS exploits).
   • DA list of all open ports on the VM and the API endpoints.
   Reveal Solution Discussion Select an answer first
2. 12expert · hard

   You are assessing a black-box translation API. You need to determine whether the model is a transformer-based model or an older recurrent neural network (RNN). You have a limited number of API calls due to rate limits. Which approach would be most efficient in distinguishing between the two?

   • ASend the same sentence multiple times to check for output variability.
   • BMeasure the API's response time for short vs. long sentences.
   • CSend multiple random sentences and compare the translation quality.
   • DCraft a single input with a long-range dependency (e.g., a sentence where the beginning determines the translation of the end) and observe if the model correctly handles it.
   Reveal Solution Discussion Select an answer first
4. 13application · medium

   You are testing a fraud detection model that operates on tabular data. You want to understand which features are most influential in the model's decisions to plan an evasion attack. You have black-box access to the model's prediction API. What is the most effective technique?

   • AAnalyze the API's documentation for details on the model's training data.
   • BRequest the model's feature importance scores from the API.
   • CPerform feature perturbation analysis by changing one feature at a time and observing the change in the prediction score.
   • DTrain a local interpretable model (e.g., logistic regression) on synthetic data to approximate feature importance.
   Reveal Solution Discussion Select an answer first
5. 14application · medium

   After mapping an AI system's attack surface, you have identified a public API, a model with known vulnerabilities, and a training pipeline that uses an untrusted dataset. Which of the following best prioritizes the attack vectors?

   • APrioritize the training pipeline because data poisoning is always the most severe.
   • BPrioritize the model because known vulnerabilities are always the most critical.
   • CPrioritize only the public API because it is the most accessible.
   • DPrioritize the public API for adversarial input attacks, the model for known exploits, and the training pipeline for data poisoning.
   Reveal Solution Discussion Select an answer first
6. 15foundation · easy

   What is the purpose of adversarial input probing in AI reconnaissance?

   • ATo test the model's accuracy on a validation dataset
   • BTo infer the model's behavior and decision boundaries
   • CTo measure the model's inference latency for performance tuning
   • DTo directly exploit a vulnerability and gain control of the model
   Reveal Solution Discussion Select an answer first