Palo Alto Networks NetSec-Architect Practice Test 13
Question 120
For an original educational Network Security Architect scenario, consider: A public generative AI application sends prompts containing trade secrets to an external model service and needs policy enforcement while requests are in flight. Select TWO complementary steps: the targeted correction and its strongest prechange validation.
- Change: Architect AI runtime protection at a supported enforcement point that can inspect prompts and responses for policy-defined risks before sensitive content is released.
- Check: Review discovered AI applications, associated users, sanctioned status, DLP event samples and enforcement rules for enterprise versus personal accounts.
- Approve every AI message simply because the endpoint uses TLS and the model previously passed a static scan.
- Change: Add a model-assessment stage using appropriate Prisma AIRS AI model scanning and approved remediation or acceptance criteria before production promotion.
- Check: Verify the actual application integration path, inspection coverage, data classification tests, blocking outcome and telemetry on permitted and rejected prompts.
Explanation
Correct Answer
A:AI-specific runtime threats can appear in model inputs, outputs and tool interactions; model scanning prior to release cannot inspect every later conversation. Acceptance: Sensitive content and malicious AI requests trigger documented controls while approved benign model interactions continue normally.
E:AI-specific runtime threats can appear in model inputs, outputs and tool interactions; model scanning prior to release cannot inspect every later conversation. Required intervention: Architect AI runtime protection at a supported enforcement point that can inspect prompts and responses for policy-defined risks before sensitive content is released.
Incorrect Answers
B:Applied case: A public generative AI application sends prompts containing trade secrets to an external model service and needs policy enforcement while requests are in flight.
C:AI-specific runtime threats can appear in model inputs, outputs and tool interactions; model scanning prior to release cannot inspect every later conversation.
D:Change: Add a model-assessment stage using appropriate Prisma AIRS AI model scanning and approved remediation or acceptance criteria before production promotion. AI-specific runtime threats can appear in model inputs, outputs and tool interactions; model scanning prior to release cannot inspect every later conversation.
Question 121
A network security architect is reviewing this environment: An internal AI purchasing agent can submit supplier payments, but its tool access has unrestricted privileges beyond the user's allowed business role. A proposed shortcut is: Grant the AI agent administrator access to every back-end system to reduce the number of failed tool calls. What should the security architect recommend as the narrowly scoped correction? Select ONE best answer.
- Combine appropriate application identification and cloud-delivered classification features with granular AI application access policy and validation.
- Architect cluster-aware workload segmentation and suitable policy or runtime inspection for inference and agent components with narrowly defined communication paths.
- Design an agent boundary with scoped tool permissions, verified user context, untrusted-content isolation and runtime controls for dangerous tool calls.
- Grant the AI agent administrator access to every back-end system to reduce the number of failed tool calls.
Explanation
Correct Answer
C:An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
Incorrect Answers
A:Combine appropriate application identification and cloud-delivered classification features with granular AI application access policy and validation. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
B:Architect cluster-aware workload segmentation and suitable policy or runtime inspection for inference and agent components with narrowly defined communication paths. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
D:An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
Question 122
A Palo Alto Networks architecture review finds: An internal AI purchasing agent can submit supplier payments, but its tool access has unrestricted privileges beyond the user's allowed business role. A design team proposes: Design an agent boundary with scoped tool permissions, verified user context, untrusted-content isolation and runtime controls for dangerous tool calls. Which architecture principle most strongly supports the recommended response? Select ONE best answer.
- An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
- Availability and risk posture must be planned together; an uninspected backup path or silent fail-open can defeat otherwise correct AI security controls.
- Kubernetes namespaces do not automatically enforce traffic isolation; security policy must account for workload identity, pod churn and network enforcement location.
- Adversarial evaluation probes model and application behavior under malicious prompts, while perimeter firewall tests alone do not expose unsafe agent reasoning or tool permissions.
Explanation
Correct Answer
A:Demonstrable success requires: Untrusted text cannot cause unauthorized tool execution, and permitted tasks retain auditable identity and change approval.
Incorrect Answers
B:Availability and risk posture must be planned together; an uninspected backup path or silent fail-open can defeat otherwise correct AI security controls. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
C:Kubernetes namespaces do not automatically enforce traffic isolation; security policy must account for workload identity, pod churn and network enforcement location. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
D:Adversarial evaluation probes model and application behavior under malicious prompts, while perimeter firewall tests alone do not expose unsafe agent reasoning or tool permissions. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
Question 123
The architecture board has not yet authorized configuration changes. Current evidence indicates: An internal AI purchasing agent can submit supplier payments, but its tool access has unrestricted privileges beyond the user's allowed business role. Which evidence best tests the architecture diagnosis before approving a disruptive change? Select ONE best answer.
- Review discovered AI applications, associated users, sanctioned status, DLP event samples and enforcement rules for enterprise versus personal accounts.
- Match change tickets to artifact hashes, release configurations, risk assessments, red-team results and actual runtime policy coverage.
- Test untrusted retrieval content, authorization context, tool invocation logs, deny decisions and the business approval flow for representative high-impact actions.
- Document personal-data classes, cross-border routes, access rights, logging and retention, control ownership and regulatory requirement mapping.
Explanation
Correct Answer
C:An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
Incorrect Answers
A:Review discovered AI applications, associated users, sanctioned status, DLP event samples and enforcement rules for enterprise versus personal accounts. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
B:Match change tickets to artifact hashes, release configurations, risk assessments, red-team results and actual runtime policy coverage. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
D:Document personal-data classes, cross-border routes, access rights, logging and retention, control ownership and regulatory requirement mapping. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
Question 124
The architecture team has implemented an approved correction and must now perform a production-safe acceptance test. Original problem: An internal AI purchasing agent can submit supplier payments, but its tool access has unrestricted privileges beyond the user's allowed business role. The operations owner requires positive and negative test evidence and confirmation that legitimate business workflows still function. Which acceptance result demonstrates improved security without breaking authorized workflows? Select ONE best answer.
- The targeted AI function is classified and handled according to business approval while unrelated authorized SaaS workflows continue.
- Every production AI release has corresponding documented assessment and controls; changed behavior is not silently approved by an older scan.
- Required inference and retrieval paths work under rescheduling, while unrelated workloads cannot access model-management or confidential dataset services.
- Untrusted text cannot cause unauthorized tool execution, and permitted tasks retain auditable identity and change approval.
Explanation
Correct Answer
D:An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
Incorrect Answers
A:The targeted AI function is classified and handled according to business approval while unrelated authorized SaaS workflows continue. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
B:Every production AI release has corresponding documented assessment and controls; changed behavior is not silently approved by an older scan. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
C:Required inference and retrieval paths work under rescheduling, while unrelated workloads cannot access model-management or confidential dataset services. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
Question 125
A design sign-off request identifies the following problem: An internal AI purchasing agent can submit supplier payments, but its tool access has unrestricted privileges beyond the user's allowed business role. Which paired security change and validation should close the design review? Select ONE best answer.
- Change: Architect cluster-aware workload segmentation and suitable policy or runtime inspection for inference and agent components with narrowly defined communication paths. Verify: Map namespace labels, pod service accounts, east-west flows, network policies and the selected enforcement integration under scale and rolling upgrades.
- Change: Design an agent boundary with scoped tool permissions, verified user context, untrusted-content isolation and runtime controls for dangerous tool calls. Verify: Test untrusted retrieval content, authorization context, tool invocation logs, deny decisions and the business approval flow for representative high-impact actions.
- Change: Design redundant AI inspection paths with explicit fail-open versus fail-closed policy, capacity and observability for every primary and recovery route. Verify: Exercise representative inspection failures and regional failover while checking request routes, policy decisions, latency and alerting.
- Change: Define gated model, prompt, tool and application changes with appropriate model scanning, adversarial retesting and updated runtime enforcement before promotion. Verify: Match change tickets to artifact hashes, release configurations, risk assessments, red-team results and actual runtime policy coverage.
Explanation
Correct Answer
B:An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements. The change is successful only when: Untrusted text cannot cause unauthorized tool execution, and permitted tasks retain auditable identity and change approval.
Incorrect Answers
A:Change: Architect cluster-aware workload segmentation and suitable policy or runtime inspection for inference and agent components with narrowly defined communication paths.
C:Change: Design redundant AI inspection paths with explicit fail-open versus fail-closed policy, capacity and observability for every primary and recovery route.
D:Change: Define gated model, prompt, tool and application changes with appropriate model scanning, adversarial retesting and updated runtime enforcement before promotion.
Question 126
A network security architect is reviewing this environment: A support automation bot is allowed to query customer records, yet injected instructions in a retrieved document cause it to attempt changes to account ownership. A proposed shortcut is: Grant the AI agent administrator access to every back-end system to reduce the number of failed tool calls. What should the security architect recommend as the narrowly scoped correction? Select ONE best answer.
- Map AI model and application data flows to applicable regulatory obligations and risk controls, then choose architecture and enforcement based on documented requirements.
- Design an agent boundary with scoped tool permissions, verified user context, untrusted-content isolation and runtime controls for dangerous tool calls.
- Add a model-assessment stage using appropriate Prisma AIRS AI model scanning and approved remediation or acceptance criteria before production promotion.
- Grant the AI agent administrator access to every back-end system to reduce the number of failed tool calls.
Explanation
Correct Answer
B:An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
Incorrect Answers
A:Map AI model and application data flows to applicable regulatory obligations and risk controls, then choose architecture and enforcement based on documented requirements. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
C:Add a model-assessment stage using appropriate Prisma AIRS AI model scanning and approved remediation or acceptance criteria before production promotion. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
D:An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
Question 127
A Palo Alto Networks architecture review finds: A support automation bot is allowed to query customer records, yet injected instructions in a retrieved document cause it to attempt changes to account ownership. A design team proposes: Design an agent boundary with scoped tool permissions, verified user context, untrusted-content isolation and runtime controls for dangerous tool calls. Which architecture principle most strongly supports the recommended response? Select ONE best answer.
- A security evaluation is version-specific; introducing new model weights, tools or data integrations can invalidate previous assurance assumptions.
- An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
- Kubernetes namespaces do not automatically enforce traffic isolation; security policy must account for workload identity, pod churn and network enforcement location.
- Adaptive application identification can improve visibility of changing AI services, but decisions still require an explicit sanctioning and least-privilege policy.
Explanation
Correct Answer
B:Demonstrable success requires: Untrusted text cannot cause unauthorized tool execution, and permitted tasks retain auditable identity and change approval.
Incorrect Answers
A:A security evaluation is version-specific; introducing new model weights, tools or data integrations can invalidate previous assurance assumptions. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
C:Kubernetes namespaces do not automatically enforce traffic isolation; security policy must account for workload identity, pod churn and network enforcement location. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
D:Adaptive application identification can improve visibility of changing AI services, but decisions still require an explicit sanctioning and least-privilege policy. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
Question 128
The architecture board has not yet authorized configuration changes. Current evidence indicates: A support automation bot is allowed to query customer records, yet injected instructions in a retrieved document cause it to attempt changes to account ownership. Which evidence best tests the architecture diagnosis before approving a disruptive change? Select ONE best answer.
- Inventory model versions, provenance, scan findings, ownership and deployment gates, then confirm that the scanned artifact matches the deployed release.
- Trace each model client, inference endpoint, control plane, identity boundary, payload path and failure behavior before selecting the integration method.
- Test synthetic sensitive records, matching classifiers, channel visibility, blocked requests, user notification and exception logging.
- Test untrusted retrieval content, authorization context, tool invocation logs, deny decisions and the business approval flow for representative high-impact actions.
Explanation
Correct Answer
D:An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements. Confirm afterward that: Untrusted text cannot cause unauthorized tool execution, and permitted tasks retain auditable identity and change approval.
Incorrect Answers
A:Inventory model versions, provenance, scan findings, ownership and deployment gates, then confirm that the scanned artifact matches the deployed release. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
B:This option addresses a different security architecture topic (AIRS inline versus API form factors) rather than the documented AI agent privilege and tool boundary decision.
C:Test synthetic sensitive records, matching classifiers, channel visibility, blocked requests, user notification and exception logging. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
Question 129
The architecture team has implemented an approved correction and must now perform a production-safe acceptance test. Original problem: A support automation bot is allowed to query customer records, yet injected instructions in a retrieved document cause it to attempt changes to account ownership. The operations owner requires positive and negative test evidence and confirmation that legitimate business workflows still function. Which acceptance result demonstrates improved security without breaking authorized workflows? Select ONE best answer.
- Untrusted text cannot cause unauthorized tool execution, and permitted tasks retain auditable identity and change approval.
- Approved enterprise AI workflows remain usable and unsanctioned data transfers are logged or prevented per policy.
- The design provides evidence of appropriate classification, access, retention and review for applicable rules without assuming a single universal control.
- Protected information is denied or masked according to classification while approved non-sensitive prompts complete normally.
Explanation
Correct Answer
A:An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements. The approved design step was: Design an agent boundary with scoped tool permissions, verified user context, untrusted-content isolation and runtime controls for dangerous tool calls.
Incorrect Answers
B:Approved enterprise AI workflows remain usable and unsanctioned data transfers are logged or prevented per policy. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
C:The design provides evidence of appropriate classification, access, retention and review for applicable rules without assuming a single universal control. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
D:Protected information is denied or masked according to classification while approved non-sensitive prompts complete normally. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
Question 130
For an original educational Network Security Architect scenario, consider: A support automation bot is allowed to query customer records, yet injected instructions in a retrieved document cause it to attempt changes to account ownership. Select TWO complementary steps: the targeted correction and its strongest prechange validation.
- Check: Examine the evaluation plan, attack prompts, guardrail failures, permission scope and retest results for the exact deployed agent release.
- Change: Design an agent boundary with scoped tool permissions, verified user context, untrusted-content isolation and runtime controls for dangerous tool calls.
- Grant the AI agent administrator access to every back-end system to reduce the number of failed tool calls.
- Change: Combine appropriate application identification and cloud-delivered classification features with granular AI application access policy and validation.
- Check: Test untrusted retrieval content, authorization context, tool invocation logs, deny decisions and the business approval flow for representative high-impact actions.
Explanation
Correct Answer
B:An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements. Acceptance: Untrusted text cannot cause unauthorized tool execution, and permitted tasks retain auditable identity and change approval.
E:An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements. Required intervention: Design an agent boundary with scoped tool permissions, verified user context, untrusted-content isolation and runtime controls for dangerous tool calls.
Incorrect Answers
A:Check: Examine the evaluation plan, attack prompts, guardrail failures, permission scope and retest results for the exact deployed agent release. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
C:An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.
D:Change: Combine appropriate application identification and cloud-delivered classification features with granular AI application access policy and validation. An LLM agent becomes a privileged actor when it invokes tools; prompt content does not grant authorization to exceed identity, application or data entitlements.