AI Reconnaissance and Attack Surface Mapping COASP Practice Questions (Page 7)

Questions 31–35

1. 31application · medium

   A client wants you to assess the external attack surface of their machine learning platform. They have provided no documentation. Which of the following is the most effective way to begin gathering publicly available information about their models and infrastructure?

   • AAttempt to access their internal Jupyter notebooks by guessing common URLs.
   • BSearch for the company's model names, dataset names, and employee GitHub accounts on public code repositories and data-sharing sites.
   • CSend a phishing email to employees to trick them into revealing internal system details.
   • DRun a vulnerability scanner against their public IP range to identify open ports and services.
   Reveal Solution Discussion Select an answer first
2. 32application · medium

   You are performing model fingerprinting on a black-box NLP API. You have access to the API but not to the model's internals. Which technique would be most effective in identifying the underlying model family (e.g., BERT vs. GPT)?

   • ACraft inputs that exploit known differences in tokenization and attention patterns, and observe the model's output behavior.
   • BMeasure the API's latency for different input lengths to estimate the model's context window size.
   • CCheck the API's rate limits to infer the model's computational complexity.
   • DSend the same prompt multiple times and analyze the randomness of the outputs.
   Reveal Solution Discussion Select an answer first
4. 33expert · hard

   During reconnaissance, you find a public S3 bucket that appears to contain model artifacts. You also find a GitHub repository with code that references the same bucket. However, the bucket listing is disabled. What is the most effective way to determine if the bucket contains sensitive model files?

   • AAttempt to enable bucket listing by exploiting a misconfiguration.
   • BMove on because the bucket is not accessible without listing.
   • CUse the GitHub repository to identify exact file names and attempt to access them directly in the bucket.
   • DTry common bucket names and paths to guess file locations.
   Reveal Solution Discussion Select an answer first
5. 34expert · hard

   You are mapping the attack surface of a microservices-based AI platform. You have discovered the main inference API, but you suspect there are internal services (e.g., model management, data preprocessing) that might be exposed through a service mesh or API gateway. You have limited time and need to prioritize. Which approach is most effective?

   • AScan the entire public IP range for open ports and services.
   • BAttempt to access the internal services directly by guessing their hostnames.
   • CFocus only on the main inference API and ignore internal services.
   • DAnalyze the API gateway's routing rules and error messages to infer internal service names and paths.
   Reveal Solution Discussion Select an answer first
6. 35foundation · easy

   Which of the following is a key step in systematically mapping an AI attack surface?

   • ACorrelating reconnaissance findings with known attack techniques
   • BTraining the AI model on a larger dataset
   • CEncrypting the model's weights at rest
   • DDeploying a firewall to protect the AI system
   Reveal Solution Discussion Select an answer first