OWASP LLM and ML Top 10 (2025) Mapping to AI Threat and Governance COASP Practice Questions (Page 9)

Questions 41–45

1. 41application · medium

   A cybersecurity team is testing a machine learning model that detects phishing websites. The team wants to evaluate the model's robustness against attackers who modify URLs to evade detection. Which OWASP ML Top 10 risk should the test focus on, and what is the appropriate offensive technique?

   • AModel Inversion; attempt to reconstruct the training data
   • BSupply Chain Attack; compromise the model's hosting environment
   • CModel Evasion; use adversarial examples that mimic URL obfuscation techniques
   • DData Poisoning; inject malicious URLs into the training set
   Reveal Solution Discussion Select an answer first
2. 42application · medium

   An offensive security team is conducting a penetration test on a web application that uses an LLM to generate responses and a separate ML model to detect malicious content. The team wants to test for OWASP LLM and ML Top 10 risks. Which testing activity is most aligned with the OWASP frameworks?

   • AAttempt to bypass the content filter using prompt injection and test the ML model with adversarial inputs.
   • BPerform a denial-of-service attack on the LLM API to test availability.
   • CTest the web application's authentication and authorization mechanisms.
   • DReview the LLM's training data for bias and fairness issues.
   Reveal Solution Discussion Select an answer first
4. 43application · medium

   A data science team has trained a fraud-detection model on historical transaction data. During testing, the model performs well on the overall dataset but fails to detect fraud for a specific demographic group. The security team is asked to assess the model's risk using the OWASP ML Top 10. Which ML risk is most directly indicated, and what is the most appropriate mitigation?

   • AAI Bias and Fairness — audit the training data for representational bias and apply fairness-aware techniques.
   • BML Model Inversion Attack — retrain the model with differential privacy.
   • CData Poisoning — sanitize the training data to remove malicious samples.
   • DMachine Learning Model Skew — implement continuous monitoring and retraining on new data.
   Reveal Solution Discussion Select an answer first
5. 44application · medium

   A security analyst is investigating a machine learning model that recommends investment portfolios. The analyst discovers that the model's recommendations are biased against certain demographics. Which OWASP ML Top 10 risk is this, and what is the primary governance concern?

   • AModel Inversion; the model is leaking sensitive information about users
   • BData Poisoning; the training data may have been tampered with
   • CModel Evasion; attackers are manipulating inputs to cause biased outputs
   • DUnfairness or Bias; the model may have learned biased patterns from training data
   Reveal Solution Discussion Select an answer first
6. 45application · medium

   An offensive AI team is planning a penetration test for a system that uses an LLM to generate reports and a separate ML model to classify documents. The team wants to test the system's resilience to attacks that could cause it to produce incorrect or harmful outputs. Which approach should the team take?

   • ATest the system's network security instead of the AI components
   • BOnly test the LLM, since it is the primary interface for users
   • COnly test the ML model, since it is more critical for classification
   • DTest both the LLM and the ML model for relevant risks, such as prompt injection and model evasion
   Reveal Solution Discussion Select an answer first