Question 800
The security team implements Agent tool permission boundaries for this requirement: A document-search agent should not obtain permissions to delete Cloud Storage buckets merely because it can search them. Which result best demonstrates successful enforcement?
Explanation
Correct Answer
This directly validates Agent tool permission boundaries: The search agent retrieves permitted documents but deletion tool calls fail authorization. The observed result is consistent with the implemented mechanism, grant only the tool-specific actions and protected service permissions needed, adding controlled review for irreversible operations, for the requirement (a document-search agent should not obtain permissions to delete cloud storage buckets merely because it can search them).
Incorrect Answers
This observation verifies AI model output data-leak testing: A user lacking source rights cannot obtain confidential text through a model answer. That validates a different security control used when an agent is able to summarize internal documents but must not reproduce a confidential document verbatim for an unauthorized user. It does not establish Agent tool permission boundaries for this scenario (a document-search agent should not obtain permissions to delete cloud storage buckets merely because it can search them).
This observation verifies Model Armor prompt injection screening: An injected source document is blocked by the configured screening policy and cannot trigger a privileged action. That validates a different security control used when a retrieval-augmented agent may ingest instructions hidden in external documents that attempt to override its system rules. It does not establish Agent tool permission boundaries for this scenario (a document-search agent should not obtain permissions to delete cloud storage buckets merely because it can search them).
This observation verifies AI training dataset access minimization: The training run uses de-identified records and cannot read raw personally identifiable information. That validates a different security control used when a model-training job should use a pseudonymized dataset rather than a raw customer identity table. It does not establish Agent tool permission boundaries for this scenario (a document-search agent should not obtain permissions to delete cloud storage buckets merely because it can search them).






Professional Cloud Security Engineer Practice Tests