AI Reconnaissance and Attack Surface Mapping COASP Practice Questions (Page 4)

Questions 16–20

1. 16application · medium

   You are testing a text classification API. You suspect it is a fine-tuned version of a well-known open-source model. Which technique would most reliably confirm the base model?

   • AAnalyze the API's error messages for framework-specific stack traces.
   • BSend inputs that are known to trigger characteristic outputs in the suspected base model and compare the API's responses.
   • CCheck the API's documentation for the model version.
   • DMeasure the API's response time for different input lengths to estimate the model's parameter count.
   Reveal Solution Discussion Select an answer first
2. 17expert · hard

   You are fingerprinting a language model API. You have a suspicion it is a specific open-source model, but the API adds random noise to the outputs to prevent exact fingerprinting. Which technique is most likely to still identify the base model?

   • ASend the same input multiple times and average the outputs to remove the noise.
   • BSend inputs that elicit long, structured responses and analyze the statistical distribution of tokens, which is less affected by random noise.
   • CUse a single, highly specific prompt that is known to produce a unique output in the suspected model.
   • DMeasure the API's response time to infer the model size.
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   You are testing a spam filter that classifies emails as spam or not. You want to infer the model's decision boundary to craft emails that bypass the filter. Which probing approach is most effective?

   • AStart with a benign email and gradually add spam-like features (e.g., certain words, formatting) until the classification flips.
   • BSend a large number of random emails and record the spam probability.
   • CSend the same email multiple times to check for classification consistency.
   • DSend emails with missing headers to see if the API returns an error.
   Reveal Solution Discussion Select an answer first
5. 19application · medium

   After completing reconnaissance on a client's AI system, you have identified: a public API endpoint, a model card indicating a TensorFlow model, and a training pipeline that uses a public dataset. Which of the following best represents a systematic correlation of these findings with potential attack vectors?

   • AConclude that the system is secure because the dataset is public and the model is well-documented.
   • BRecommend immediate patching of the TensorFlow framework without further analysis.
   • CMap the public API to potential adversarial input attacks, the TensorFlow model to known framework-specific exploits, and the public dataset to data poisoning risks.
   • DFocus only on the public API because it is the only externally accessible component.
   Reveal Solution Discussion Select an answer first
6. 20foundation · easy

   Which of the following is an example of adversarial input probing?

   • ASending inputs with slight perturbations to observe changes in the model's output
   • BIntercepting and modifying network traffic to the API
   • CUsing SQL injection to extract data from the backend database
   • DSending a large number of requests to overload the server
   Reveal Solution Discussion Select an answer first