Cisco 810-110 Practice Test 27
Question 260
A public dataset is modified to associate a trusted phrase with a malicious response. If the relevant Training data poisoning safeguard is omitted, which outcome is the MOST directly associated failure?
- Installing invented package names can execute attacker-controlled code.
- A source document may make a harmless summarization task perform exfiltration.
- Blindly mixing attacker-controlled records can compromise model responses.
- Assuming a local hostname guarantees all provider processing is local can cause violations.
Explanation
Correct Answer
C:Blindly mixing attacker-controlled records can compromise model responses. Corrupted training examples can steer learned behavior or embed backdoors. The corrective approach is to use provenance checks, curation, and held-out adversarial validation.
Incorrect Answers
A:Installing invented package names can execute attacker-controlled code. This failure pattern belongs chiefly to AI-generated dependency security risks; it is not the failure most directly tested in the Training data poisoning situation. Blindly mixing attacker-controlled records can compromise model responses.
B:A source document may make a harmless summarization task perform exfiltration. This failure pattern belongs chiefly to Tool-enabled prompt injection security; it is not the failure most directly tested in the Training data poisoning situation. Blindly mixing attacker-controlled records can compromise model responses.
D:Assuming a local hostname guarantees all provider processing is local can cause violations. This failure pattern belongs chiefly to Geographic data residency approval; it is not the failure most directly tested in the Training data poisoning situation. Blindly mixing attacker-controlled records can compromise model responses.
Question 261
A mobile assistant must give a two-sentence troubleshooting hint to fit the user interface. After the recommended Constraining permitted answer length change is implemented, which observation MOST directly verifies that it works?
- Output reliably fits the interface without losing the required action.
- Independent calculations and test outputs match the model's final claim.
- The explanation is understandable while no unauthorized operations are performed.
- Adversarial tests do not override the application's protected handling rules.
Explanation
Correct Answer
A:Output reliably fits the interface without losing the required action. This directly tests whether the Constraining permitted answer length goal was met. Output format and length are explicit prompt constraints rather than implicit preferences.
Incorrect Answers
B:Independent calculations and test outputs match the model's final claim. This is evidence about External calculator and code validation, not direct proof of the required Constraining permitted answer length behavior. The relevant check is that output reliably fits the interface without losing the required action.
C:The explanation is understandable while no unauthorized operations are performed. This is evidence about Persona tailoring without authority escalation, not direct proof of the required Constraining permitted answer length behavior. The relevant check is that output reliably fits the interface without losing the required action.
D:Adversarial tests do not override the application's protected handling rules. This is evidence about System versus user instruction priorities, not direct proof of the required Constraining permitted answer length behavior. The relevant check is that output reliably fits the interface without losing the required action.
Question 262
A client contract requires information to stay within a named processing region. Which practical correction or implementation decision is MOST directly appropriate?
- Select the lowest-priced overseas endpoint without checking residency obligations.
- Implement server-side role checks and minimal tool credentials.
- Audit complete processing paths and select contractually approved regions.
- Redact and pseudonymize data at the boundary before inference.
Explanation
Correct Answer
C:Audit complete processing paths and select contractually approved regions. The intended behavior can be checked because data-flow inventories and contracts show processing stays within approved jurisdictions.
Incorrect Answers
A:Select the lowest-priced overseas endpoint without checking residency obligations. This would not correct the Geographic data residency approval workflow. Assuming a local hostname guarantees all provider processing is local can cause violations.
B:Implement server-side role checks and minimal tool credentials. This could help Role-based access for AI service tools, but it does not resolve the Geographic data residency approval case. Residency can include telemetry, storage, and backups, not only the primary API endpoint.
D:Redact and pseudonymize data at the boundary before inference. This could help Personal data redaction before prompting, but it does not resolve the Geographic data residency approval case. Residency can include telemetry, storage, and backups, not only the primary API endpoint.
Question 263
A technician wants an AI assistant to interpret a photograph of a hardware connector. Which approach BEST meets the requirement while respecting the system constraints?
- Choose a suitable text-to-image diffusion model with rights and safety checks.
- Send only the JPEG filename to a text-only model and expect reliable visual perception.
- Select an instruction-tuned language model that predicts and generates text.
- Use a multimodal vision-language model that accepts images plus text.
Explanation
Correct Answer
D:Use a multimodal vision-language model that accepts images plus text. Vision-language models jointly process image and language inputs for visual reasoning tasks. Verification should demonstrate that the answer identifies visible relevant features and avoids unsupported hidden details.
Incorrect Answers
A:Choose a suitable text-to-image diffusion model with rights and safety checks. This primarily addresses Diffusion-based image creation, not the specific Multimodal visual question answering decision. Here, vision-language models jointly process image and language inputs for visual reasoning tasks.
B:Send only the JPEG filename to a text-only model and expect reliable visual perception. This is an unsafe or ineffective shortcut for Multimodal visual question answering. An image-blind language model may invent what a photo contains without receiving pixels.
C:Select an instruction-tuned language model that predicts and generates text. This primarily addresses Autoregressive language generation versus image diffusion, not the specific Multimodal visual question answering decision. Here, vision-language models jointly process image and language inputs for visual reasoning tasks.
Question 264
A department rolls out AI tools without an escalation route for harmful outputs. If the relevant Policy enforcement and incident response safeguard is omitted, which outcome is the MOST directly associated failure?
- Automatically publishing generated compliance claims can create false representations.
- A one-size-fits-all approval path may underprotect high-stakes uses.
- Leaving embedding stores and transcripts unencrypted can disclose sensitive relationships.
- Unreported model incidents can persist without correction or customer notice.
Explanation
Correct Answer
D:Unreported model incidents can persist without correction or customer notice. Governance requires operation-time policies and remediation, not only launch approval. The corrective approach is to train owners and establish incident reporting and containment.
Incorrect Answers
A:Automatically publishing generated compliance claims can create false representations. This failure pattern belongs chiefly to Documented human approval for regulated changes; it is not the failure most directly tested in the Policy enforcement and incident response situation. Unreported model incidents can persist without correction or customer notice.
B:A one-size-fits-all approval path may underprotect high-stakes uses. This failure pattern belongs chiefly to AI use-case risk classification; it is not the failure most directly tested in the Policy enforcement and incident response situation. Unreported model incidents can persist without correction or customer notice.
C:Leaving embedding stores and transcripts unencrypted can disclose sensitive relationships. This failure pattern belongs chiefly to Encryption and key management of AI data; it is not the failure most directly tested in the Policy enforcement and incident response situation. Unreported model incidents can persist without correction or customer notice.
Question 265
An organization must prevent fabricated media from being treated as verified evidence. Select TWO statements that correctly combine the Synthetic misinformation and impersonation method with evidence or its governing principle.
- Require corroborating evidence and accountable human judgment for material decisions.
- Validate dataset provenance, scan anomalies, and isolate untrusted training contributions.
- Verify provenance with independent sources and flag synthetic or uncertain media.
- Editorial checks distinguish authenticated footage from manipulated claims.
- Assume photorealism proves a video is an authentic recording.
Explanation
Correct Answers
C:Verify provenance with independent sources and flag synthetic or uncertain media. This is the intended action for Synthetic misinformation and impersonation.
D:Editorial checks distinguish authenticated footage from manipulated claims. This is directly observable validation of the Synthetic misinformation and impersonation result.
Incorrect Answers
A:Require corroborating evidence and accountable human judgment for material decisions. This serves the different Automation bias and overreliance objective, not the selected Synthetic misinformation and impersonation plan. Generative AI can produce convincing but false text, images, and audio.
B:Validate dataset provenance, scan anomalies, and isolate untrusted training contributions. This serves the different Training data poisoning objective, not the selected Synthetic misinformation and impersonation plan. Generative AI can produce convincing but false text, images, and audio.
E:Assume photorealism proves a video is an authentic recording. This is not part of a safe or effective Synthetic misinformation and impersonation design. Publishing unverified generated content can amplify misinformation and reputational harm.
Question 266
A searchable knowledge base needs to match paraphrases like password reset and account recovery. After the recommended Embedding similarity retrieval change is implemented, which observation MOST directly verifies that it works?
- Data flow traces show scoped snippets and no unrelated restricted records leaving the boundary.
- The final context contains the most relevant passages on a judged test set.
- New versions appear and deleted versions no longer surface in queries.
- Paraphrased queries retrieve relevant indexed passages in held-out tests.
Explanation
Correct Answer
D:Paraphrased queries retrieve relevant indexed passages in held-out tests. This directly tests whether the Embedding similarity retrieval goal was met. Similarity search compares embedding geometry; it does not guarantee factual truth.
Incorrect Answers
A:Data flow traces show scoped snippets and no unrelated restricted records leaving the boundary. This is evidence about Hybrid AI architecture with private retrieval, not direct proof of the required Embedding similarity retrieval behavior. The relevant check is that paraphrased queries retrieve relevant indexed passages in held-out tests.
B:The final context contains the most relevant passages on a judged test set. This is evidence about Reranking retrieved passages, not direct proof of the required Embedding similarity retrieval behavior. The relevant check is that paraphrased queries retrieve relevant indexed passages in held-out tests.
C:New versions appear and deleted versions no longer surface in queries. This is evidence about Retrieval index freshness, not direct proof of the required Embedding similarity retrieval behavior. The relevant check is that paraphrased queries retrieve relevant indexed passages in held-out tests.
Question 267
An assistant must synthesize structured highlights from multiple documents with traceability. Which practical correction or implementation decision is MOST directly appropriate?
- Preserve chunk provenance through staged synthesis and cross-check coverage.
- Improve relevance selection and run position-robustness tests.
- Combine explicit output controls with completeness validation and retry policy.
- Concatenate every unbounded document and hope the model remembers each one.
Explanation
Correct Answer
A:Preserve chunk provenance through staged synthesis and cross-check coverage. The intended behavior can be checked because the merged summary cites source chunks and covers all required sections.
Incorrect Answers
B:Improve relevance selection and run position-robustness tests. This could help Lost-in-the-middle retrieval quality, but it does not resolve the Long-document chunk and merge case. Map-reduce summarization trades extra steps for bounded context and source traceability.
C:Combine explicit output controls with completeness validation and retry policy. This could help Stop conditions and maximum output, but it does not resolve the Long-document chunk and merge case. Map-reduce summarization trades extra steps for bounded context and source traceability.
D:Concatenate every unbounded document and hope the model remembers each one. This would not correct the Long-document chunk and merge workflow. Dropping source identifiers during aggregation makes claims hard to verify.
Question 268
An internal AI assistant can search documents and invoke production APIs. Which approach BEST meets the requirement while respecting the system constraints?
- Give the assistant a shared root key and rely on a polite system prompt.
- Use TLS in transit, encryption at rest, and controlled key access for stored artifacts.
- Grant narrowly scoped tools and roles based on authenticated user permissions.
- Select compliant processing regions and verify backup, log, and subprocessors' locations.
Explanation
Correct Answer
C:Grant narrowly scoped tools and roles based on authenticated user permissions. Least-privilege authorization must be enforced outside natural-language instructions. Verification should demonstrate that attempted unauthorized tool calls are denied and logged by backend enforcement.
Incorrect Answers
A:Give the assistant a shared root key and rely on a polite system prompt. This is an unsafe or ineffective shortcut for Role-based access for AI service tools. A model with blanket administrator credentials can turn one prompt error into a breach.
B:Use TLS in transit, encryption at rest, and controlled key access for stored artifacts. This primarily addresses Encryption and key management of AI data, not the specific Role-based access for AI service tools decision. Here, least-privilege authorization must be enforced outside natural-language instructions.
D:Select compliant processing regions and verify backup, log, and subprocessors' locations. This primarily addresses Geographic data residency approval, not the specific Role-based access for AI service tools decision. Here, least-privilege authorization must be enforced outside natural-language instructions.
Question 269
A regulated European business is evaluating inference endpoints in several jurisdictions. If the relevant Geographic data residency approval safeguard is omitted, which outcome is the MOST directly associated failure?
- Blindly mixing attacker-controlled records can compromise model responses.
- Assuming a local hostname guarantees all provider processing is local can cause violations.
- Copying raw customer records into third-party prompts may violate policy.
- Leaving embedding stores and transcripts unencrypted can disclose sensitive relationships.
Explanation
Correct Answer
B:Assuming a local hostname guarantees all provider processing is local can cause violations. Residency can include telemetry, storage, and backups, not only the primary API endpoint. The corrective approach is to audit complete processing paths and select contractually approved regions.
Incorrect Answers
A:Blindly mixing attacker-controlled records can compromise model responses. This failure pattern belongs chiefly to Training data poisoning; it is not the failure most directly tested in the Geographic data residency approval situation. Assuming a local hostname guarantees all provider processing is local can cause violations.
C:Copying raw customer records into third-party prompts may violate policy. This failure pattern belongs chiefly to Personal data redaction before prompting; it is not the failure most directly tested in the Geographic data residency approval situation. Assuming a local hostname guarantees all provider processing is local can cause violations.
D:Leaving embedding stores and transcripts unencrypted can disclose sensitive relationships. This failure pattern belongs chiefly to Encryption and key management of AI data; it is not the failure most directly tested in the Geographic data residency approval situation. Assuming a local hostname guarantees all provider processing is local can cause violations.
Question 270
A document-processing agent recursively calls an expensive model while retrying failures. Which approach BEST meets the requirement while respecting the system constraints?
- Increase retry concurrency without limit whenever an API reports throttling.
- Set concurrency, retry, token, and spending budgets with circuit breakers.
- Define explicit states, branch conditions, and terminal outcomes with logs.
- Design compensating actions or resumable checkpoints around side effects.
Explanation
Correct Answer
B:Set concurrency, retry, token, and spending budgets with circuit breakers. AI workflow capacity needs explicit budget and backpressure controls. Verification should demonstrate that run metrics show bounded model usage without exhausting service quotas.
Incorrect Answers
A:Increase retry concurrency without limit whenever an API reports throttling. This is an unsafe or ineffective shortcut for Rate limits and cost budgets. Unbounded loops can produce runaway spend and cascading throttling.
C:Define explicit states, branch conditions, and terminal outcomes with logs. This primarily addresses Multi-step workflow branching, not the specific Rate limits and cost budgets decision. Here, aI workflow capacity needs explicit budget and backpressure controls.
D:Design compensating actions or resumable checkpoints around side effects. This primarily addresses Workflow compensation after partial failure, not the specific Rate limits and cost budgets decision. Here, aI workflow capacity needs explicit budget and backpressure controls.