Question 290
A company collects chats to resolve support cases. A team proposes reusing them to infer employee performance without an approved purpose or data review. What should happen first?
- Assess the proposed new purpose, permissions, and applicable governance requirements
- Train the model first and seek approval only if results are useful
- Proceed because internal ownership permits any use of collected data
- Remove customer names and assume every secondary use is permitted
Explanation
Correct Answer
A:A new use can change expectations, risks, and obligations even when the data already exists. The organization must establish whether that secondary use is authorized.
Incorrect Answers
B:Model development would already use the data for the unapproved purpose. Review should precede that processing rather than depend on its apparent usefulness.
C:Possession alone does not establish an unrestricted purpose. The proposed employee analysis requires its own justification and governance assessment.
D:Removing names may reduce some risk but does not authorize a new purpose. Remaining content may also retain sensitive or identifying information.
Question 291
An application's retention policy requires support content to expire after 30 days. The main database deletes it, but traces and backups keep accessible copies indefinitely. What is missing?
- A shorter generated summary for each support conversation
- A retention design covering each copy, including backup lifecycle and access restrictions
- A stronger instruction telling the model to forget old conversations
- A report counting successful deletions from the primary table only
Explanation
Correct Answer
B:The requirement concerns the data throughout its lifecycle, not one table. Traces and backups need defined handling consistent with the approved retention design.
Incorrect Answers
A:Shorter summaries may contain less content but do not establish expiration. The indefinite copies still require lifecycle controls.
C:The model's instructions do not delete stored traces or backups. Retention must be implemented in the systems that hold the copies.
D:That count measures one storage location and omits the known retained copies. It cannot demonstrate compliance with the application's full retention requirement.
Question 292
An analytics assistant suppresses names but returns a salary average for a group containing one person. What privacy issue remains?
- The query is safe whenever the underlying records were encrypted at rest
- The average becomes confidential only if the assistant prints the person's name
- The query is safe because averages cannot equal individual records
- The aggregate directly reveals an individual value despite removing the name
Explanation
Correct Answer
D:Aggregation does not protect confidentiality when the group is small enough to isolate a person. The output policy must consider identifiability and query context, not only explicit names.
Incorrect Answers
A:Storage encryption does not prevent disclosure through an authorized computation's output. The issue is what the returned statistic reveals to the requester.
B:A user may know who belongs to the one-person group. Identity can be inferred from context even when the generated output omits a name.
C:For a group of one, the average equals the sole value. The mathematical form of an aggregate does not guarantee privacy.
Question 293
A contract restricts specified customer content to Region A. Inference runs there, but an error-reporting service exports raw prompts to Region B. What is the correct assessment?
- The requirement applies only to successful requests
- The complete processing path must be reviewed because telemetry crosses the stated boundary
- The export is acceptable if the raw prompts are deleted after error diagnosis
- The export is acceptable whenever the telemetry service encrypts it
- The requirement is satisfied because the model runs in Region A
Explanation
Correct Answer
B:Residency applies to the specified content wherever it is processed or retained under the contract. Regional inference alone does not cover an out-of-region telemetry export.
Incorrect Answers
A:Errors can contain the same restricted content as successful requests. Request outcome does not remove the contract's stated data restriction.
C:Shorter retention does not undo the transfer across the stated regional boundary. The processing path must satisfy the contract when data moves.
D:Encryption protects confidentiality but does not change the destination region. It cannot by itself satisfy a restriction on where specified content may go.
E:The prompt content also reaches the error service. Focusing only on inference ignores a known processing location outside the stated boundary.
Question 294
A health organization asks whether signing a business associate agreement alone proves its proposed AI workflow meets all applicable HIPAA requirements. Which response is appropriate?
- Yes, if the assistant includes a medical disclaimer in every response
- Assess the agreement alongside deployment-specific roles and safeguards
- No, because all AI processing of health information is categorically prohibited
- Yes, the agreement automatically certifies every connected application
Explanation
Correct Answer
B:An agreement does not replace deployment-specific responsibilities and safeguards. The organization must evaluate how the proposed workflow handles the information in practice.
Incorrect Answers
A:A disclaimer does not implement privacy or security safeguards. It cannot establish compliance for processing protected information.
C:A blanket prohibition does not follow from the scenario. The relevant decision depends on applicable requirements, agreements, roles, and implemented safeguards.
D:A contract does not validate all application configurations or third-party components. The broader workflow still has responsibilities that require assessment.
Question 295
A public-sector team uses a cloud service offering with FedRAMP Certified status. It plans to claim the same status for its entire custom assistant. What should the architect verify?
- The authorized boundary and coverage of the proposed deployment
- Whether the team uses the same cloud region as another government customer
- Only that the component appears somewhere in the architecture diagram
- Whether the application reuses the component's compliance logo
Explanation
Correct Answer
A:FedRAMP status applies to an assessed cloud service offering and its scope. A service's status should not be generalized to an unassessed custom application assembled around it.
Incorrect Answers
B:Another customer's region does not establish this system's authorization scope. The actual deployment and its applicable boundary need verification.
C:Using an authorized component does not automatically extend its status to every connected service. Boundary and deployment coverage must be established.
D:A logo does not establish which services, configurations, and responsibilities are covered. The proposed system needs evidence tied to its actual authorization boundary.
Question 296
A privacy team requires a process for responding to applicable access and correction requests. The AI application has scattered stores and no responsible owner. What is the most useful architectural contribution?
- Map relevant records and derived stores, then assign a coordinated request-handling workflow
- Offer a universal legal interpretation directly from the system prompt
- Promise that the model will remember every user's corrections permanently
- Delete all application data immediately whenever any request arrives
Explanation
Correct Answer
A:An actionable process needs to locate relevant information and coordinate responsible owners. Architecture can make the required workflow traceable across stores and services.
Incorrect Answers
B:A prompt is not an authoritative determination of each request's legal scope. The architecture should support the qualified privacy team's process.
C:A model response is not a reliable record-management workflow. Corrections must reach the appropriate stores under the organization's approved process.
D:Applicable requests may require verification, scoping, and different actions. An indiscriminate deletion rule does not implement the stated access and correction workflow.
Question 297
A vendor states that API inputs are not used to train its models under the relevant terms. A stakeholder concludes that inputs are therefore never retained anywhere. What clarification is needed?
- No clarification is needed because storage always implies training
- Training use and retention are separate questions requiring separate evidence
- The retention period can be inferred from the maximum supported context window
- Deleting the local chat view necessarily deletes every provider copy
- The conclusion follows whenever the provider returns no conversation identifier
Explanation
Correct Answer
B:A restriction on training use does not by itself specify storage duration or every processing purpose. Retention must be checked against the relevant service terms and configuration.
Incorrect Answers
A:Storage can support several purposes unrelated to training. Treating the two as identical leads to an unsupported retention conclusion.
C:Context capacity describes how much input a request can use, not how provider systems store it. Retention requires separate policy or contractual evidence.
D:A local interface action does not establish deletion across external services. The relevant deletion and retention behavior must be verified separately.
E:The absence of a client-visible identifier does not establish absence of storage. Backend retention practices need direct evidence.
Question 298
A project has a contractual zero-retention arrangement for eligible endpoints. It adds a new feature whose eligibility has not been confirmed. What should block production use of that feature with covered data?
- The fact that the feature uses the same provider brand
- The absence of a retention setting in the feature's interface
- The inability to disable optional analytics in the local user interface alone
- The unresolved retention applicability for the specific feature and deployment path
Explanation
Correct Answer
D:An arrangement may have scope and feature conditions. The new path needs confirmed eligibility before the team relies on the existing retention assumption.
Incorrect Answers
A:A shared provider name does not prove identical data handling across features. Eligibility must be checked at the relevant service boundary.
B:An absent control does not establish the underlying data-handling terms. Eligibility must be confirmed for the specific feature and deployment path.
C:A local setting may affect one data path but does not establish contractual eligibility for a new feature. The specific service path and applicable terms need confirmation.
Question 299
An incident reveals that a live API credential was written into an accessible trace. Which immediate action most directly removes the credential's continuing utility to an attacker?
- Add a prompt instructing the assistant not to repeat the credential
- Delete the trace but leave the credential active
- Revoke or rotate the exposed credential and update legitimate consumers
- Wait until unauthorized usage appears in billing reports
Explanation
Correct Answer
C:An exposed credential can remain usable after the trace is hidden. Revocation or rotation removes that continuing access while affected services receive replacement credentials.
Incorrect Answers
A:The model's future wording does not invalidate the exposed secret. Credential validity must be changed at the issuing system.
B:Deleting one visible copy does not establish that nobody retained the secret. The active credential could still be used by anyone who obtained it.
D:Waiting preserves an avoidable access path. Usage evidence can support investigation, but it is not necessary to revoke a known exposed credential.
Question 300
An autonomous workflow begins issuing harmful external updates. Its normal approval service is also unavailable. Which previously implemented control would provide the most dependable containment?
- An approval request sent through the unavailable approval service
- An independently operable switch that disables the workflow's write capability
- A dashboard that displays the number of harmful updates
- An instruction in the workflow asking it to stop after the current task
Explanation
Correct Answer
B:Containment must remain available when an ordinary dependency fails. Disabling writes outside the affected workflow prevents further actions while responders investigate.
Incorrect Answers
A:That path depends on the component already known to be unavailable. It cannot provide dependable immediate containment in this scenario.
C:Visibility supports response but does not stop the actions. The team needs an enforceable control that can disable further writes.
D:The affected workflow may not reliably honor that instruction. It also allows the current harmful task to continue before containment.