Question 90
Original unofficial Palo Alto Networks NetSec-Architect scenario: A financial institution wants to test whether a planned assistant follows malicious instructions embedded inside retrieved internal documents before deployment. Which test result best demonstrates that the design meets its requirement?
Explanation
Correct Answer
Original fictional security architecture situation: A financial institution wants to test whether a planned assistant follows malicious instructions embedded inside retrieved internal documents before deployment. Governing principle: Adversarial red teaming exposes model-and-workflow behavior that normal functional tests may miss. The most defensible design response is: Include an authorized AI red-teaming assessment of indirect prompt-injection cases in the preproduction launch gate. Validation: Controlled injected documents fail to induce prohibited disclosure or unauthorized tool actions in the candidate assistant. Sustained assurance: Rerun adversarial scenarios after major retrieval, model, and tool changes
Incorrect Answers
The suggested “Test prompts to the external API are inspected according to the approved runtime rules.” tests a different architectural concern: differentiate prisma airs and ai access capabilities. Here, a financial institution wants to test whether a planned assistant follows malicious instructions embedded inside retrieved internal documents before deployment. Adversarial red teaming exposes model-and-workflow behavior that normal functional tests may miss.
“Both flows have explicit inspection and audit evidence without requiring identical infrastructure components.” concerns the separate design topic: select ai security reference architectures, form factors, and data controls. It would not fix the described weakness: a financial institution wants to test whether a planned assistant follows malicious instructions embedded inside retrieved internal documents before deployment. Adversarial red teaming exposes model-and-workflow behavior that normal functional tests may miss.
This alternative is linked to the architecture category: select ai security reference architectures, form factors, and data controls. The actual problem is that a financial institution wants to test whether a planned assistant follows malicious instructions embedded inside retrieved internal documents before deployment; the controlling distinction is: adversarial red teaming exposes model-and-workflow behavior that normal functional tests may miss.