AI Vulnerability Scanning and Fuzzing COASP Practice Questions (Page 7)

Questions 31–35

1. 31foundation · easy

   Which of the following is a common framework used for adversarial robustness testing of AI models?

   • ANmap
   • BBurp Suite
   • CCleverHans
   • DMetasploit
   Reveal Solution Discussion Select an answer first
2. 32application · medium

   After running a fuzzing campaign on a chatbot model, the security team observes that the model outputs offensive language when given inputs containing certain Unicode characters. The team needs to decide whether this is a security vulnerability and how to prioritize it. What is the most appropriate interpretation and action?

   • AIgnore it because fuzzing often produces unrealistic inputs that will not occur in production.
   • BClassify it as an AI-specific vulnerability (e.g., a safety failure) and implement input filtering or output moderation as an immediate mitigation.
   • CReport it as a traditional injection vulnerability and patch the web application firewall.
   • DTreat it as a low-priority quality issue and defer any action until the next model update.
   Reveal Solution Discussion Select an answer first
4. 33application · medium

   A vulnerability scan of a facial recognition system reveals that the model is highly susceptible to adversarial perturbations that cause misidentification. The system is used for physical access control. Which mitigation strategy would be most effective in reducing the risk while maintaining system usability?

   • AIncrease the model's confidence threshold to reduce false accepts.
   • BRetrain the model with adversarial training and add input preprocessing that reduces the impact of perturbations.
   • CDisable the facial recognition system and require manual ID checks.
   • DAdd a rate limit to the API to slow down attackers.
   Reveal Solution Discussion Select an answer first
5. 34foundation · easy

   What is a recommended mitigation for data poisoning vulnerabilities?

   • ADisabling the model's logging
   • BUsing a larger batch size during training
   • CImplementing strict data validation and anomaly detection on the training dataset
   • DIncreasing the model's inference timeout
   Reveal Solution Discussion Select an answer first
6. 35foundation · easy

   Which AI vulnerability category is characterized by an attacker crafting small, imperceptible changes to input data to cause misclassification?

   • AAdversarial perturbations
   • BModel inversion
   • CData poisoning
   • DMembership inference
   Reveal Solution Discussion Select an answer first