Question 20
An engineer chooses AI model output data-leak testing because an agent is able to summarize internal documents but must not reproduce a confidential document verbatim for an unauthorized user. Which technical description of the selected control is accurate?
Explanation
Correct Answer
The operative mechanism for AI model output data-leak testing: Enforce retrieval authorization and output data-protection checks before presenting model responses. For this scenario (an agent is able to summarize internal documents but must not reproduce a confidential document verbatim for an unauthorized user), verify that a user lacking source rights cannot obtain confidential text through a model answer.
Incorrect Answers
This describes how Agent tool permission boundaries works instead: Grant only the tool-specific actions and protected service permissions needed, adding controlled review for irreversible operations. It is useful when a purchase-order assistant must require authorized approval before creating irreversible financial transactions, but it does not explain the selected AI model output data-leak testing for the current requirement (an agent is able to summarize internal documents but must not reproduce a confidential document verbatim for an unauthorized user).
This describes how AI training service-account separation works instead: Execute training jobs under dedicated identities with the minimum approved dataset and storage access. It is useful when a fine-tuning pipeline must not inherit the notebook author's broad credentials when launched unattended, but it does not explain the selected AI model output data-leak testing for the current requirement (an agent is able to summarize internal documents but must not reproduce a confidential document verbatim for an unauthorized user).
This describes how Gemini Agent Platform Agent Identity works instead: Use Agent Identity and appropriately scoped policy grants for each agent workload and tool-access path. It is useful when a support agent should not inherit broad project editor permissions simply because its developer has them, but it does not explain the selected AI model output data-leak testing for the current requirement (an agent is able to summarize internal documents but must not reproduce a confidential document verbatim for an unauthorized user).






Professional Cloud Security Engineer Practice Tests