OWASP LLM and ML Top 10 (2025) Mapping to AI Threat and Governance COASP Practice Questions (Page 2)

Questions 6–10

1. 6expert · hard

   A company is deploying an LLM-based system that automatically responds to customer complaints on social media. The system has access to customer data and can post replies. The legal team is concerned about the system making defamatory statements, and the security team is concerned about prompt injection. Which governance control should be implemented to address both concerns?

   • ARequire human approval for all posts and implement input filtering to block prompt injection attempts
   • BDisable the system's ability to post replies, only allowing it to draft responses
   • CUse a more advanced LLM that is less likely to generate defamatory content
   • DImplement a content filter that blocks any post containing negative words
   Reveal Solution Discussion Select an answer first
2. 7foundation · easy

   An attacker sends a specially crafted message to an LLM-based chatbot that causes it to ignore its system prompt and reveal hidden instructions. Which OWASP LLM Top 10 risk does this scenario map to?

   • ALLM03: Supply Chain
   • BLLM02: Sensitive Information Disclosure
   • CLLM01: Prompt Injection
   • DLLM04: Insecure Output Handling
   Reveal Solution Discussion Select an answer first
4. 8expert · hard

   A security team is assessing a machine learning model that recommends products to users. The team has white-box access to the model and its training data. They want to test for the risk of model theft. Which approach is most effective?

   • AUse the model's API to send queries and observe the outputs to infer the model's architecture.
   • BExtract the model's parameters by querying the model and using the training data to train a substitute model.
   • CPerform a membership inference attack to determine if specific records were in the training data.
   • DCheck the model's accuracy on a test set to see if it is overfitting.
   Reveal Solution Discussion Select an answer first
5. 9foundation · easy

   Which OWASP ML Top 10 risk is characterized by an attacker causing the model to make incorrect predictions by manipulating its inputs?

   • AML06: Membership Inference
   • BML08: Data Poisoning
   • CML07: Adversarial Perturbation
   • DML05: Model Inversion
   Reveal Solution Discussion Select an answer first
6. 10expert · hard

   A security team is assessing a machine learning model that classifies email attachments as malicious or benign. The team has a limited budget and must choose between testing for model evasion or model poisoning. The model is retrained weekly on user-reported emails. Which risk should be prioritized, and why?

   • AModel poisoning, because the retraining pipeline is a direct attack surface for injecting malicious samples
   • BModel evasion, because attackers will always try to craft emails that bypass the classifier
   • CSupply chain attack, because the model may be hosted on a third-party platform
   • DModel inversion, because the model may leak sensitive information about training data
   Reveal Solution Discussion Select an answer first