Question 420
A prompt-update pipeline already validates syntax, evaluates thresholds, and requires release-owner approval. An incident shows that a change passed those gates but was never exercised against realistic test queries during CI or checked after deployment. Which TWO automated activities add those missing pre-release and post-release checks? Choose TWO.
- Prompt syntax validation
- Threshold-based pipeline failure
- Post-release smoke test
- Automated evaluation job
- Deployment approval
Explanation
Correct Answers
C:Smoke tests catch wiring and credential defects after promotion; they complement deeper offline evaluations but cannot establish comprehensive quality. Here, the new endpoint is checked after promotion rather than relying solely on pipeline success.
D:Evaluation jobs generate comparable quality results for a proposed revision; they should use pinned model and dataset versions to limit confounding changes. Here, a stable representative test set exercises the changed behavior before release.
Incorrect Answers
A:Static validation catches deterministic template errors early; it cannot assess whether a well-formed prompt produces correct answers; it does not by itself address representative CI behavioral evaluation and a postdeployment functional check.
B:An enforced threshold prevents accidental promotion of known regressions; a dashboard warning without pipeline enforcement is weaker release protection; the observed fault still involves representative CI behavioral evaluation and a postdeployment functional check.
E:Approvals create accountability for accepted residual risk; a passing automated test does not eliminate the need for domain-specific oversight; this mechanism leaves unresolved representative CI behavioral evaluation and a postdeployment functional check.
Question 421
An enterprise agent retrieves untrusted documents and can invoke approved business tools. The immediate design goal is to treat retrieved passages as data even when they contain imperatives that resemble higher-priority instructions. Which approach addresses that goal?
- Per-tool authorization
- Untrusted context boundary
- Injection regression tests
- Sensitive action confirmation
Explanation
Correct Answer
B:Separating context from authority reduces prompt injection risk. Documents should inform answers without granting themselves permission to redefine agent policy.
Incorrect Answers
A:Tool authorization prevents an agent from exceeding the user's allowed access; model compliance alone is not a reliable security boundary; the requirement, by contrast, is to treat retrieved passages as data even when they contain imperatives that resemble higher-priority instructions.
C:Repeatable attack tests reveal whether changes reopen prior vulnerabilities; they should include both direct user and indirect document-borne attempts; the specified outcome instead requires the team to treat retrieved passages as data even when they contain imperatives that resemble higher-priority instructions.
D:Confirmation adds a control point before actions such as deleting records or changing permissions; it cannot be replaced by the model claiming its reasoning was confident; the implementation under review needs to treat retrieved passages as data even when they contain imperatives that resemble higher-priority instructions.
Question 422
An enterprise agent retrieves untrusted documents and can invoke approved business tools. The team is required to expose only the permitted tool operations for a task instead of giving an agent every available administrative function. Which option is the best fit?
- Untrusted context boundary
- Sensitive action confirmation
- Injection regression tests
- Tool allowlist
Explanation
Correct Answer
D:A tool allowlist limits what the agent can execute even when prompted maliciously. It is an authorization control separate from output content moderation.
Incorrect Answers
A:Separating context from authority reduces prompt injection risk; documents should inform answers without granting themselves permission to redefine agent policy; the implementation under review needs to expose only the permitted tool operations for a task instead of giving an agent every available administrative function.
B:Confirmation adds a control point before actions such as deleting records or changing permissions; it cannot be replaced by the model claiming its reasoning was confident; the requested task remains to expose only the permitted tool operations for a task instead of giving an agent every available administrative function.
C:Repeatable attack tests reveal whether changes reopen prior vulnerabilities; they should include both direct user and indirect document-borne attempts; this alternative does not achieve the requirement to expose only the permitted tool operations for a task instead of giving an agent every available administrative function.
Question 423
An enterprise agent retrieves untrusted documents and can invoke approved business tools. An operator is tasked to check the initiating user's identity and permissions when an agent requests access to protected business records. Which mechanism provides that capability?
- Per-tool authorization
- Untrusted context boundary
- Injection regression tests
- Tool allowlist
Explanation
Correct Answer
A:Tool authorization prevents an agent from exceeding the user's allowed access. Model compliance alone is not a reliable security boundary.
Incorrect Answers
B:Separating context from authority reduces prompt injection risk; documents should inform answers without granting themselves permission to redefine agent policy; the outstanding need is to check the initiating user's identity and permissions when an agent requests access to protected business records.
C:Repeatable attack tests reveal whether changes reopen prior vulnerabilities; they should include both direct user and indirect document-borne attempts; this approach leaves the team unable to check the initiating user's identity and permissions when an agent requests access to protected business records.
D:A tool allowlist limits what the agent can execute even when prompted maliciously; it is an authorization control separate from output content moderation; the needed capability is to check the initiating user's identity and permissions when an agent requests access to protected business records.
Question 424
An enterprise agent retrieves untrusted documents and can invoke approved business tools. The immediate design goal is to strip or quarantine unsafe markup and executable content from retrieved sources before passing them into processing tools. Which approach addresses that goal?
- Sensitive action confirmation
- External content sanitization
- Per-tool authorization
- Untrusted context boundary
Explanation
Correct Answer
B:Content handling reduces the chance that untrusted documents exploit downstream parsers or interfaces. It does not guarantee a prompt cannot be manipulated.
Incorrect Answers
A:Confirmation adds a control point before actions such as deleting records or changing permissions; it cannot be replaced by the model claiming its reasoning was confident; the needed capability is to strip or quarantine unsafe markup and executable content from retrieved sources before passing them into processing tools.
C:Tool authorization prevents an agent from exceeding the user's allowed access; model compliance alone is not a reliable security boundary; the team would still need to strip or quarantine unsafe markup and executable content from retrieved sources before passing them into processing tools.
D:Separating context from authority reduces prompt injection risk; documents should inform answers without granting themselves permission to redefine agent policy; the specified outcome instead requires the team to strip or quarantine unsafe markup and executable content from retrieved sources before passing them into processing tools.
Question 425
An enterprise agent retrieves untrusted documents and can invoke approved business tools. The team is required to maintain adversarial retrieval examples that request secrets or tool misuse as part of every release evaluation. Which option is the best fit?
- Injection regression tests
- External content sanitization
- Per-tool authorization
- Tool allowlist
Explanation
Correct Answer
A:Repeatable attack tests reveal whether changes reopen prior vulnerabilities. They should include both direct user and indirect document-borne attempts.
Incorrect Answers
B:Content handling reduces the chance that untrusted documents exploit downstream parsers or interfaces; it does not guarantee a prompt cannot be manipulated; the implementation under review needs to maintain adversarial retrieval examples that request secrets or tool misuse as part of every release evaluation.
C:Tool authorization prevents an agent from exceeding the user's allowed access; model compliance alone is not a reliable security boundary; the requested task remains to maintain adversarial retrieval examples that request secrets or tool misuse as part of every release evaluation.
D:A tool allowlist limits what the agent can execute even when prompted maliciously; it is an authorization control separate from output content moderation; this alternative does not achieve the requirement to maintain adversarial retrieval examples that request secrets or tool misuse as part of every release evaluation.
Question 426
An agent processes untrusted supplier HTML and retrieved documents that may contain fake 'system' instructions, and it can initiate high-impact purchases. Tool allowlisting and caller authorization are already enforced. Which THREE additional runtime safeguards prevent external text from taking authority, prevent unsafe markup from reaching tools, and require trusted approval for a purchase? Choose THREE.
- Per-tool authorization
- Tool allowlist
- External content sanitization
- Sensitive action confirmation
- Injection regression tests
- Untrusted context boundary
Explanation
Correct Answers
C:Content handling reduces the chance that untrusted documents exploit downstream parsers or interfaces; it does not guarantee a prompt cannot be manipulated. Here, hostile active content is removed or isolated before it can reach downstream processors.
D:Confirmation adds a control point before actions such as deleting records or changing permissions; it cannot be replaced by the model claiming its reasoning was confident. Here, an irreversible action waits for an authorized confirmation outside the untrusted document.
F:Separating context from authority reduces prompt injection risk; documents should inform answers without granting themselves permission to redefine agent policy. Here, document text is never upgraded into a trusted system or developer instruction.
Incorrect Answers
A:Tool authorization prevents an agent from exceeding the user's allowed access; model compliance alone is not a reliable security boundary; the observed fault still involves untrusted text authority, risky external markup, and high-impact action approval.
B:A tool allowlist limits what the agent can execute even when prompted maliciously; it is an authorization control separate from output content moderation; the immediate requirement instead concerns untrusted text authority, risky external markup, and high-impact action approval.
E:Repeatable attack tests reveal whether changes reopen prior vulnerabilities; they should include both direct user and indirect document-borne attempts; it does not by itself address untrusted text authority, risky external markup, and high-impact action approval.
Question 427
A live Foundry-based agent needs operational controls during degraded service events. The immediate design goal is to handle a sudden increase in rate-limit responses with controlled retries and traffic reduction before modifying prompts. Which approach addresses that goal?
- Incident trace correlation
- Model throttling runbook
- Agent tool timeout fallback
- Provider outage fallback
- Emergency prompt rollback
Explanation
Correct Answer
B:A throttling playbook addresses capacity pressure through backoff and admission control. It should not misclassify every HTTP 429 as incorrect model behavior.
Incorrect Answers
A:Distributed traces reconstruct the sequence of dependent calls during an incident; an isolated model latency average cannot reveal which tool failed; this approach leaves the team unable to handle a sudden increase in rate-limit responses with controlled retries and traffic reduction before modifying prompts.
C:Timeout handling prevents indefinite agent execution and unsupported invented results; a robust fallback distinguishes unavailable evidence from a successful tool result; the requirement, by contrast, is to handle a sudden increase in rate-limit responses with controlled retries and traffic reduction before modifying prompts.
D:Fallback routing can improve continuity when the backup meets security and quality requirements; unvalidated model substitution risks different output contracts and behaviors; the team would still need to handle a sudden increase in rate-limit responses with controlled retries and traffic reduction before modifying prompts.
E:Prompt rollback restores an approved behavior baseline without waiting for model retraining; it requires immutable release references and a functioning deployment workflow; the specified outcome instead requires the team to handle a sudden increase in rate-limit responses with controlled retries and traffic reduction before modifying prompts.
Question 428
A live Foundry-based agent needs operational controls during degraded service events. The team is required to route eligible low-risk requests to a pretested alternate deployment when the primary serving path is unavailable. Which option is the best fit?
- Emergency prompt rollback
- Incident trace correlation
- Agent tool timeout fallback
- Provider outage fallback
Explanation
Correct Answer
D:Fallback routing can improve continuity when the backup meets security and quality requirements. Unvalidated model substitution risks different output contracts and behaviors.
Incorrect Answers
A:Prompt rollback restores an approved behavior baseline without waiting for model retraining; it requires immutable release references and a functioning deployment workflow; the team would still need to route eligible low-risk requests to a pretested alternate deployment when the primary serving path is unavailable.
B:Distributed traces reconstruct the sequence of dependent calls during an incident; an isolated model latency average cannot reveal which tool failed; the specified outcome instead requires the team to route eligible low-risk requests to a pretested alternate deployment when the primary serving path is unavailable.
C:Timeout handling prevents indefinite agent execution and unsupported invented results; a robust fallback distinguishes unavailable evidence from a successful tool result; the implementation under review needs to route eligible low-risk requests to a pretested alternate deployment when the primary serving path is unavailable.
Question 429
A live Foundry-based agent needs operational controls during degraded service events. An operator is tasked to return a safe partial response or escalate when a required external tool does not answer before its deadline. Which mechanism provides that capability?
- Provider outage fallback
- Agent tool timeout fallback
- Model throttling runbook
- Incident trace correlation
Explanation
Correct Answer
B:Timeout handling prevents indefinite agent execution and unsupported invented results. A robust fallback distinguishes unavailable evidence from a successful tool result.
Incorrect Answers
A:Fallback routing can improve continuity when the backup meets security and quality requirements; unvalidated model substitution risks different output contracts and behaviors; the requested task remains to return a safe partial response or escalate when a required external tool does not answer before its deadline.
C:A throttling playbook addresses capacity pressure through backoff and admission control; it should not misclassify every HTTP 429 as incorrect model behavior; the actual design criterion is to return a safe partial response or escalate when a required external tool does not answer before its deadline.
D:Distributed traces reconstruct the sequence of dependent calls during an incident; an isolated model latency average cannot reveal which tool failed; the outstanding need is to return a safe partial response or escalate when a required external tool does not answer before its deadline.
Question 430
A live Foundry-based agent needs operational controls during degraded service events. The immediate design goal is to link an end-user error to its model request and downstream tool spans using a shared correlation identifier. Which approach addresses that goal?
- Model throttling runbook
- Emergency prompt rollback
- Incident trace correlation
- Postincident evaluation update
Explanation
Correct Answer
C:Distributed traces reconstruct the sequence of dependent calls during an incident. An isolated model latency average cannot reveal which tool failed.
Incorrect Answers
A:A throttling playbook addresses capacity pressure through backoff and admission control; it should not misclassify every HTTP 429 as incorrect model behavior; the outstanding need is to link an end-user error to its model request and downstream tool spans using a shared correlation identifier.
B:Prompt rollback restores an approved behavior baseline without waiting for model retraining; it requires immutable release references and a functioning deployment workflow; this approach leaves the team unable to link an end-user error to its model request and downstream tool spans using a shared correlation identifier.
D:A regression suite encodes incident lessons as repeatable quality checks; a written incident report alone will not prevent the same behavior from returning; the requirement, by contrast, is to link an end-user error to its model request and downstream tool spans using a shared correlation identifier.