Cisco 810-110 Practice Test 38
Question 370
A team needs a proposed SQL transformation and must establish whether it works. Select TWO statements that correctly combine the Code-generation model with execution tests method with evidence or its governing principle.
- Merge any plausible-looking code without running a single test.
- Use a multimodal vision-language model that accepts images plus text.
- Use a code-capable generative model, then execute tests and review the result.
- The generated program passes representative unit tests and security review.
- Encode queries and documents with a compatible embedding model and compare vector similarity.
Explanation
Correct Answers
C:Use a code-capable generative model, then execute tests and review the result. This is the intended action for Code-generation model with execution tests.
D:The generated program passes representative unit tests and security review. This is directly observable validation of the Code-generation model with execution tests result.
Incorrect Answers
A:Merge any plausible-looking code without running a single test. This is not part of a safe or effective Code-generation model with execution tests design. Deploying untested generated code can introduce runtime and security defects.
B:Use a multimodal vision-language model that accepts images plus text. This serves the different Multimodal visual question answering objective, not the selected Code-generation model with execution tests plan. Code generation proposes source text but does not prove the code executes correctly.
E:Encode queries and documents with a compatible embedding model and compare vector similarity. This serves the different Text embeddings for semantic matching objective, not the selected Code-generation model with execution tests plan. Code generation proposes source text but does not prove the code executes correctly.
Question 371
An operations team needs to know model latency, tool errors, and completion rates. Select TWO statements that correctly combine the Observability for AI workflows method with evidence or its governing principle.
- Instrument per-step traces, safe logs, quality metrics, and failure alerts.
- Operators can isolate the failing stage without leaking secret payloads.
- Pause side-effecting actions for the designated approver and record the decision.
- Record only a generic success message for every workflow outcome.
- Define explicit states, branch conditions, and terminal outcomes with logs.
Explanation
Correct Answers
A:Instrument per-step traces, safe logs, quality metrics, and failure alerts. This is the intended action for Observability for AI workflows.
B:Operators can isolate the failing stage without leaking secret payloads. This is directly observable validation of the Observability for AI workflows result.
Incorrect Answers
C:Pause side-effecting actions for the designated approver and record the decision. This serves the different Human approval gates for actions objective, not the selected Observability for AI workflows plan. Observable workflows need stage-level telemetry beyond a final text answer.
D:Record only a generic success message for every workflow outcome. This is not part of a safe or effective Observability for AI workflows design. Without trace links, intermittent failures are hard to diagnose.
E:Define explicit states, branch conditions, and terminal outcomes with logs. This serves the different Multi-step workflow branching objective, not the selected Observability for AI workflows plan. Observable workflows need stage-level telemetry beyond a final text answer.
Question 372
A security analyst uses a model summary to close an incident without reviewing original evidence. Which practical correction or implementation decision is MOST directly appropriate?
- Consider an AI-generated conclusion more authoritative than signed system logs.
- Limit sensitive training content and test leakage with governed access.
- Train reviewers to verify generated claims against source systems.
- Resolve dependencies against verified registries and review AI code.
Explanation
Correct Answer
C:Train reviewers to verify generated claims against source systems. The intended behavior can be checked because audited reviews cite primary telemetry rather than treating model prose as proof.
Incorrect Answers
A:Consider an AI-generated conclusion more authoritative than signed system logs. This would not correct the Automation bias and overreliance workflow. Uncritical acceptance can bury serious security or operational faults.
B:Limit sensitive training content and test leakage with governed access. This could help Sensitive data memorization exposure, but it does not resolve the Automation bias and overreliance case. Fluent AI outputs can induce excessive trust even when they are incorrect.
D:Resolve dependencies against verified registries and review AI code. This could help AI-generated dependency security risks, but it does not resolve the Automation bias and overreliance case. Fluent AI outputs can induce excessive trust even when they are incorrect.
Question 373
A workflow can issue account credits that require an authorized operator. An engineer must explain the recommendation for Human approval before financial side effects. Which underlying principle is the MOST accurate?
- Human-in-the-loop controls reserve consequential decisions for responsible people.
- Schema mapping translates between service contracts and preserves meaning.
- Human oversight needs evidence of both permission and actual execution.
- Human control requires meaningful intervention, not merely passive observation.
Explanation
Correct Answer
A:Human-in-the-loop controls reserve consequential decisions for responsible people. Therefore the implementation decision is to pause at an enforced human approval gate before executing the financial action.
Incorrect Answers
B:Automatically executing a model-suggested refund can exceed delegated authority. This describes the separate Mapping agent tool output to a target schema mechanism; the present case instead requires Human approval before financial side effects.
C:Automatically executing a model-suggested refund can exceed delegated authority. Human oversight needs evidence of both permission and actual execution.
D:Automatically executing a model-suggested refund can exceed delegated authority. This describes the separate Human override and reversible control mechanism; the present case instead requires Human approval before financial side effects.
Question 374
A security analyst uses a model summary to close an incident without reviewing original evidence. Select TWO statements that correctly combine the Automation bias and overreliance method with evidence or its governing principle.
- Verify dependency names, versions, source integrity, and static security findings.
- Require corroborating evidence and accountable human judgment for material decisions.
- Audited reviews cite primary telemetry rather than treating model prose as proof.
- Enforce tool authorization independent of retrieved text and sandbox unsafe actions.
- Consider an AI-generated conclusion more authoritative than signed system logs.
Explanation
Correct Answers
B:Require corroborating evidence and accountable human judgment for material decisions. This is the intended action for Automation bias and overreliance.
C:Audited reviews cite primary telemetry rather than treating model prose as proof. This is directly observable validation of the Automation bias and overreliance result.
Incorrect Answers
A:Verify dependency names, versions, source integrity, and static security findings. This serves the different AI-generated dependency security risks objective, not the selected Automation bias and overreliance plan. Fluent AI outputs can induce excessive trust even when they are incorrect.
D:Enforce tool authorization independent of retrieved text and sandbox unsafe actions. This serves the different Tool-enabled prompt injection security objective, not the selected Automation bias and overreliance plan. Fluent AI outputs can induce excessive trust even when they are incorrect.
E:Consider an AI-generated conclusion more authoritative than signed system logs. This is not part of a safe or effective Automation bias and overreliance design. Uncritical acceptance can bury serious security or operational faults.
Question 375
A team needs to compare two candidates' robustness before exposing users. An engineer must explain the recommendation for Safety and reliability evaluation of models. Which underlying principle is the MOST accurate?
- Model selection includes risk testing, not just fluent sample generation.
- Model repositories list capabilities that must match task modality requirements.
- Availability in a repository does not mean unrestricted commercial rights.
- Diffusion generation iteratively denoises a sampled representation conditioned on guidance.
Explanation
Correct Answer
A:Model selection includes risk testing, not just fluent sample generation. Therefore the implementation decision is to evaluate answer quality, refusals, prompt injection response, data leakage, and reliability on test cases.
Incorrect Answers
B:Model repositories list capabilities that must match task modality requirements. This describes the separate Modality-aware model repository selection mechanism; the present case instead requires Safety and reliability evaluation of models. A polished single demo hides harmful error modes and prompt-injection failures.
C:Availability in a repository does not mean unrestricted commercial rights. This describes the separate Model license and usage restrictions mechanism; the present case instead requires Safety and reliability evaluation of models. A polished single demo hides harmful error modes and prompt-injection failures.
D:Diffusion generation iteratively denoises a sampled representation conditioned on guidance. This describes the separate Diffusion-based image creation mechanism; the present case instead requires Safety and reliability evaluation of models. A polished single demo hides harmful error modes and prompt-injection failures.
Question 376
A generated training assistant serves users with varying visual and reading needs. After the recommended Inclusive accessibility testing change is implemented, which observation MOST directly verifies that it works?
- Usability checks show equivalent access and comprehensible information.
- High-risk inputs trigger appropriate controls and qualified review.
- Poisoned samples are detected or excluded before training and evaluation.
- Decision logs and escalation contacts identify accountable people.
Explanation
Correct Answer
A:Usability checks show equivalent access and comprehensible information. This directly tests whether the Inclusive accessibility testing goal was met. Responsible AI includes equitable access and assessment beyond average users.
Incorrect Answers
B:High-risk inputs trigger appropriate controls and qualified review. This is evidence about AI safety review for risky use cases, not direct proof of the required Inclusive accessibility testing behavior. The relevant check is that usability checks show equivalent access and comprehensible information.
C:Poisoned samples are detected or excluded before training and evaluation. This is evidence about Training data poisoning, not direct proof of the required Inclusive accessibility testing behavior. The relevant check is that usability checks show equivalent access and comprehensible information.
D:Decision logs and escalation contacts identify accountable people. This is evidence about Accountable model decision ownership, not direct proof of the required Inclusive accessibility testing behavior. The relevant check is that usability checks show equivalent access and comprehensible information.
Question 377
An analyst wants AI help with logs that include emails and private credentials. Testing exposes this failure: Copying raw customer records into third-party prompts may violate policy. Which remediation should be applied FIRST?
- Select compliant processing regions and verify backup, log, and subprocessors' locations.
- Send full customer identifiers and hope the model ignores them.
- Define approved retention, deletion, audit access, and data-use settings contractually.
- Redact and pseudonymize data at the boundary before inference.
Explanation
Correct Answer
D:Redact and pseudonymize data at the boundary before inference. Confirmation requires seeing that test prompts and telemetry contain no raw protected fields beyond approved need.
Incorrect Answers
A:Select compliant processing regions and verify backup, log, and subprocessors' locations. It addresses Geographic data residency approval, not the diagnosed Personal data redaction before prompting failure. Copying raw customer records into third-party prompts may violate policy.
B:Send full customer identifiers and hope the model ignores them. This retains the Personal data redaction before prompting problem. Data minimization reduces exposure of unnecessary information to model services.
C:Define approved retention, deletion, audit access, and data-use settings contractually. It addresses Controlled conversation retention, not the diagnosed Personal data redaction before prompting failure. Copying raw customer records into third-party prompts may violate policy.
Question 378
An agent reports that it edited a document but its update API returned an error. Which approach BEST meets the requirement while respecting the system constraints?
- Define completion criteria, maximum steps, budget ceilings, and escalation rules.
- Use the model's statement done as the only write-verification signal.
- Use RAG for evidence-backed answering and agent coordination only when actions and planning are needed.
- Verify the actual tool response and target state before reporting action completion.
Explanation
Correct Answer
D:Verify the actual tool response and target state before reporting action completion. Agent outputs should reflect confirmed external state rather than imagined progress. Verification should demonstrate that readback proves the intended change exists or the system reports failure.
Incorrect Answers
A:Define completion criteria, maximum steps, budget ceilings, and escalation rules. This primarily addresses Bounded autonomy versus unsupervised loops, not the specific Agent feedback and task verification decision. Here, agent outputs should reflect confirmed external state rather than imagined progress.
B:Use the model's statement done as the only write-verification signal. This is an unsafe or ineffective shortcut for Agent feedback and task verification. Claiming success after a failed tool invocation deceives users and masks recovery needs.
C:Use RAG for evidence-backed answering and agent coordination only when actions and planning are needed. This primarily addresses Retrieval assistant versus planning agent, not the specific Agent feedback and task verification decision. Here, agent outputs should reflect confirmed external state rather than imagined progress.
Question 379
A research agent makes dozens of unnecessary search calls when source quality is low. After the recommended Step and budget limits on agents change is implemented, which observation MOST directly verifies that it works?
- Runs stop at defined ceilings with traceable errors and partial findings.
- Every receiving agent gets verified inputs and accountable completion criteria.
- Test cases only enter states permitted by their predecessor conditions.
- A diagram correctly maps each server through its client to the coordinating host.
Explanation
Correct Answer
A:Runs stop at defined ceilings with traceable errors and partial findings. This directly tests whether the Step and budget limits on agents goal was met. Bounded execution prevents runaway autonomous loops and unexpected cost.
Incorrect Answers
B:Every receiving agent gets verified inputs and accountable completion criteria. This is evidence about Multi-agent handoff with scoped roles, not direct proof of the required Step and budget limits on agents behavior. The relevant check is that runs stop at defined ceilings with traceable errors and partial findings.
C:Test cases only enter states permitted by their predecessor conditions. This is evidence about Finite-state agent orchestration, not direct proof of the required Step and budget limits on agents behavior. The relevant check is that runs stop at defined ceilings with traceable errors and partial findings.
D:A diagram correctly maps each server through its client to the coordinating host. This is evidence about MCP host-client-server roles, not direct proof of the required Step and budget limits on agents behavior. The relevant check is that runs stop at defined ceilings with traceable errors and partial findings.
Question 380
A field worker provides a screen capture and asks for a natural-language description. An engineer must explain the recommendation for Multimodal visual question answering. Which underlying principle is the MOST accurate?
- Hybrid architectures can separate controlled data stores from inference but must enforce egress policy.
- Vision-language models jointly process image and language inputs for visual reasoning tasks.
- Code generation proposes source text but does not prove the code executes correctly.
- Autoregressive LLMs generate token sequences whereas diffusion models commonly synthesize images by denoising.
Explanation
Correct Answer
B:Vision-language models jointly process image and language inputs for visual reasoning tasks. Therefore the implementation decision is to use a multimodal vision-language model that accepts images plus text.
Incorrect Answers
A:An image-blind language model may invent what a photo contains without receiving pixels. Hybrid architectures can separate controlled data stores from inference but must enforce egress policy.
C:An image-blind language model may invent what a photo contains without receiving pixels. Code generation proposes source text but does not prove the code executes correctly.
D:An image-blind language model may invent what a photo contains without receiving pixels. This describes the separate Autoregressive language generation versus image diffusion mechanism; the present case instead requires Multimodal visual question answering.