OWASP LLM and ML Top 10 (2025) Mapping to AI Threat and Governance COASP Practice Questions (Page 3)

Questions 11–15

1. 11expert · hard

   An offensive AI team is planning a red-team exercise for a system that uses both an LLM for customer support and a separate ML model for fraud detection. The team has limited time and must choose which system to test first. The LLM has access to customer PII, and the ML model blocks transactions. Which factor should drive the prioritization?

   • ATest the LLM first because it is easier to attack
   • BThe ML model should be tested first because it directly affects financial transactions
   • CTest both simultaneously to save time, even if the assessment is less thorough
   • DThe LLM should be tested first because it has access to sensitive PII, and a breach would have higher regulatory impact
   Reveal Solution Discussion Select an answer first
2. 12application · medium

   A healthcare organization is deploying an LLM-based clinical decision-support system. The compliance team requires that the system not leak patient data and that any model output be traceable to its source. Which governance policy should be implemented to address the OWASP LLM Top 10 risk of Sensitive Information Disclosure?

   • ARequire the LLM to be hosted on-premises to ensure data residency
   • BRequire model versioning and rollback procedures for all updates
   • CRequire output filtering and data-loss-prevention (DLP) controls on all LLM responses
   • DRequire all LLM prompts to be logged for audit purposes
   Reveal Solution Discussion Select an answer first
4. 13expert · hard

   A multinational corporation is deploying an LLM-based HR assistant that handles employee inquiries. The company operates in multiple countries with different data protection regulations. The security team must design a governance framework that addresses OWASP LLM Top 10 risks while complying with regional data residency requirements. Which approach best balances these concerns?

   • ADeploy the LLM in a single region and apply the strictest data protection controls globally.
   • BUse a multi-region deployment with data residency controls and implement LLM-specific security controls such as input filtering and output validation.
   • CUse a public LLM API and rely on the provider's compliance certifications.
   • DDisable the LLM's ability to store any data and use it only for real-time responses.
   Reveal Solution Discussion Select an answer first
5. 14foundation · easy

   During an offensive AI security assessment, a tester wants to evaluate whether an LLM-based system can be tricked into revealing its system prompt. Which OWASP LLM Top 10 risk should the tester focus on?

   • ALLM05: Excessive Agency
   • BLLM02: Sensitive Information Disclosure
   • CLLM03: Supply Chain
   • DLLM01: Prompt Injection
   Reveal Solution Discussion Select an answer first
6. 15expert · hard

   An offensive AI team is conducting a red-team exercise on a system that uses an LLM to moderate user-generated content and a separate ML model to detect spam. The team has discovered that the LLM can be tricked into approving malicious content, and the ML model can be evaded with certain spam patterns. The team has limited time to report findings. Which finding should be reported as the highest risk?

   • ABoth findings are equally severe and should be reported together
   • BThe ML model evasion is more severe because it affects a larger volume of content
   • CThe LLM can be tricked into approving malicious content, which could lead to harmful content being published
   • DThe ML model can be evaded with certain spam patterns, which could lead to spam reaching users
   Reveal Solution Discussion Select an answer first