API Reconnaissance and Model Fingerprinting COASP Practice Questions (Page 7)

Questions 31–35

1. 31application · medium

   A security auditor is reviewing an AI API that returns a confidence score along with the prediction. The auditor notices that the response includes a 'debug_info' field that contains the model's feature names and a partial training-data sample. The auditor wants to determine if this is a data leakage vulnerability. Which action is most appropriate?

   • AReview the API documentation to see if the 'debug_info' field is intended to be public.
   • BReport the 'debug_info' field as a data leakage vulnerability because it exposes model metadata and training data remnants.
   • CAttempt to access the 'debug_info' field with different user roles to see if it is restricted by authorization controls.
   • DCheck if the 'debug_info' field is present in the production environment or only in a staging environment.
   Reveal Solution Discussion Select an answer first
2. 32foundation · easy

   Which of the following is a common technique to bypass API rate limiting?

   • AUsing HTTP/2 instead of HTTP/1.1
   • BRotating IP addresses or using proxies
   • CIncreasing the request payload size
   • DAdding extra headers to the request
   Reveal Solution Discussion Select an answer first
4. 33application · medium

   You are testing an image recognition API that classifies images into categories. You want to generate adversarial examples to test the model's robustness. The API has a rate limit and requires authentication. Which approach is most effective for generating adversarial examples?

   • AUse a black-box attack that queries the API thousands of times to estimate gradients.
   • BQuery the API with every possible image in the training set to find misclassifications.
   • CSend random noise to the API and observe misclassifications.
   • DUse a white-box attack on a local surrogate model and then transfer the adversarial examples to the API.
   Reveal Solution Discussion Select an answer first
5. 34application · medium

   You are performing reconnaissance on a web application that uses an AI API. You have access to the web app's JavaScript files and the API documentation. Which method is most effective for discovering undocumented API endpoints?

   • ACheck the API documentation for any hidden endpoints by looking at the source code of the documentation page.
   • BAnalyze the JavaScript files for hardcoded URLs and API paths, and compare them with the documented endpoints.
   • CUse a wordlist to brute-force common API paths on the server.
   • DSend a request to the root API path and observe the response headers for hints.
   Reveal Solution Discussion Select an answer first
6. 35application · medium

   You are testing a natural language processing API that classifies text sentiment. The API accepts a 'text' parameter and returns a confidence score. You suspect the underlying model is a transformer-based architecture. Which payload would best help you confirm the model architecture while also testing for injection vulnerabilities?

   • ASend a prompt that asks the model to reveal its own architecture in the response.
   • BSend a request with a negative sentiment text and check if the confidence score changes as expected.
   • CSend a long sequence of repeated tokens and measure the response time and output behavior to infer model context length.
   • DSend a specially crafted string with SQL injection syntax and observe if the API returns a database error.
   Reveal Solution Discussion Select an answer first