AI Vulnerability Scanning and Fuzzing COASP Practice Questions (Page 9)

Questions 41–45

1. 41expert · hard

   After a fuzzing campaign on a credit scoring model, the results show that the model is vulnerable to both data poisoning and adversarial perturbations. The team has limited resources and must decide which to remediate first. The model is retrained monthly with new data, and the inference endpoint is publicly accessible. Which remediation should be prioritized?

   • AFocus on data validation and sanitization to prevent poisoning, since the model is retrained monthly.
   • BDeploy a web application firewall to block all suspicious requests.
   • CTake the model offline until both issues are fully resolved.
   • DImplement adversarial training to make the model robust to perturbations, since the endpoint is public and attackers can easily craft inputs.
   Reveal Solution Discussion Select an answer first
2. 42application · medium

   A security analyst is planning to fuzz a deep learning model that processes tabular data. The analyst wants to use a specialized framework that can generate adversarial examples and also measure coverage of the model's neurons. Which tool or framework is most appropriate for this task?

   • AA library like TensorFuzz or DeepXplore
   • BA general-purpose HTTP fuzzer like Burp Suite
   • CTensorFlow's built-in debugger
   • DA standard unit testing framework like pytest
   Reveal Solution Discussion Select an answer first
4. 43foundation · easy

   In mutation-based fuzzing for AI, what is typically mutated?

   • AThe model's weights
   • BThe input samples
   • CThe training algorithm
   • DThe inference server's configuration
   Reveal Solution Discussion Select an answer first
5. 44foundation · easy

   Which of the following is a unique attack surface specific to AI systems?

   • AThe web server's TLS configuration
   • BThe model's training data pipeline
   • CThe database's SQL query optimization
   • DThe operating system's patch level
   Reveal Solution Discussion Select an answer first
6. 45application · medium

   A security engineer is tasked with scanning a production AI system that serves a fraud-detection model. The system includes a batch training pipeline, a feature store, a REST inference endpoint, and a model registry. The engineer wants to identify vulnerabilities that are unique to the AI components rather than the underlying web application. Which scanning approach best targets the AI-specific attack surface?

   • ARun a standard web vulnerability scanner against the REST endpoint and patch any OWASP Top 10 issues found.
   • BReview the training code for common software bugs and dependency vulnerabilities.
   • CConduct a network penetration test on the model registry and feature store databases.
   • DPerform a model-level assessment by sending crafted inputs to the inference endpoint and analyzing the feature store for poisoning indicators.
   Reveal Solution Discussion Select an answer first