AI Reconnaissance and Attack Surface Mapping COASP Practice Questions (Page 5)

Questions 21–25

1. 21application · medium

   During a red-team engagement, you discover a public API endpoint that accepts image inputs and returns a confidence score. You want to determine whether the underlying model is a ResNet-50 or a Vision Transformer (ViT) without accessing any internal documentation. Which approach is most effective?

   • ASubmit the same image multiple times and check if the confidence scores are deterministic.
   • BQuery the API with a large batch of random images and measure the average inference latency to estimate the model's depth.
   • CSend a series of carefully crafted adversarial perturbations and compare the model's misclassification patterns to known behaviors of each architecture.
   • DInspect the HTTP response headers for framework-specific version strings.
   Reveal Solution Discussion Select an answer first
2. 22application · medium

   A client's AI model is deployed behind a web application. You are asked to gather information about the model's training data and architecture without directly interacting with the model. Which source is most likely to provide this information?

   • AThe web application's robots.txt file, which may list disallowed paths to training data.
   • BThe HTTP response headers, which may include a custom header with the model version.
   • CThe SSL certificate of the web server, which may include the model's framework in the organization name.
   • DThe model card published on the company's website, which often lists training data sources, model architecture, and intended use.
   Reveal Solution Discussion Select an answer first
4. 23foundation · easy

   Which of the following is NOT typically considered a component of an AI attack surface?

   • AThe training pipeline and data preprocessing steps
   • BThe deployment infrastructure, such as servers and APIs
   • CThe model's architecture and parameters
   • DThe organization's internal communication channels
   Reveal Solution Discussion Select an answer first
5. 24application · medium

   During an assessment, you discover that an AI service exposes multiple endpoints, but only one is documented. You suspect there are hidden endpoints for model management or debugging. Which approach is most likely to reveal these hidden endpoints?

   • ASearch the web for the API's Swagger/OpenAPI specification, which may list all endpoints.
   • BSend a POST request to the documented endpoint with malformed JSON to trigger an error that reveals other routes.
   • CAnalyze the API's response time for different requests to infer the existence of hidden endpoints.
   • DUse a wordlist of common endpoint names (e.g., /admin, /debug, /internal) and fuzz the API base URL.
   Reveal Solution Discussion Select an answer first
6. 25foundation · easy

   Which of the following is a technique for collecting publicly available information about an AI system?

   • ASending crafted inputs to the model to observe outputs
   • BIntercepting network traffic between the user and the AI service
   • CSearching for model artifacts on public repositories like Hugging Face
   • DExploiting a vulnerability in the model's API to gain unauthorized access
   Reveal Solution Discussion Select an answer first