PMLE - Monitoring AI Solutions (13% of the exam) - Section 6.1

Identify and mitigate risks to AI solutions, building secure systems against data exfiltration, malicious prompting, and sensitive-data leakage using tools such as regular expressions, safety filters, and Model Armor, aligning with responsible AI practices including bias monitoring, and applying model explainability on the Agent Platform.

Identify AI risks such as data exfiltration, malicious prompting, and sensitive-data leakage, and apply mitigations including safety filters, regular expressions, and Model Armor to harden generative AI deployments. Align with responsible AI practices by monitoring for bias, and use Agent Platform model explainability tools to interpret predictions and support accountability.

Model ArmorSafety filtersResponsible AIModel explainability

Practice question for this objective

Free sampleMonitoring AI Solutionsmedium

A logistics company runs a Gemini-backed support agent on the Agent Platform. The agent can call an internal tool that returns full shipment records, and a security review shows that a crafted user message can make the agent paste a customer's saved card number and email straight into its reply. Card numbers follow a fixed 16-digit pattern, but the team also worries about novel social-engineering prompts that try to coax the model into leaking data in ways no fixed pattern anticipates. Which TWO controls should the team combine to reduce sensitive-data leakage in the model's responses? (Select TWO.)

ALower the model's temperature setting so that generated responses become more deterministic and therefore less likely to include any sensitive customer values.
BApply a regular-expression filter to the generated output to detect and redact strings matching the fixed 16-digit card-number format before the reply reaches the user. Correct
CIncrease the model's context window so it can hold the full shipment record and reason more carefully about which fields are safe to return to the user.
DGrant the agent's service account broader read access to the shipment database so it can verify records faster and avoid returning stale or partial customer data.
EPlace Model Armor on the response path to screen generated output for sensitive-data leakage and prompt-driven exfiltration patterns that no single fixed regex anticipates. Correct

Combine a deterministic regex for fixed-format values with managed Model Armor screening to stop sensitive-data leakage in model responses. Layered output controls work because a regular expression cheaply and reliably catches structured values like a 16-digit card number, while Model Armor adds managed detection of sensitive-data leakage and exfiltration patterns that no fixed pattern can anticipate; the distractors change model behaviour or access without screening output.

Why A is wrong: Tempting because lower temperature feels like a safety lever, but temperature only changes output randomness and does nothing to detect or strip sensitive data, so leakage continues.

Why B is correct: Correct because a regex reliably catches values with a known fixed structure such as a 16-digit card number, redacting them from the response cheaply and deterministically.

Why C is wrong: Tempting since a larger context seems to give the model more awareness, but window size does not enforce any data-protection rule and the model can still echo card numbers verbatim.

Why D is wrong: Tempting as a reliability tweak, but widening data access increases exfiltration exposure rather than reducing leakage and addresses none of the security concern.

Why E is correct: Correct because Model Armor is a managed control that inspects model responses for sensitive-data leakage and novel threat patterns, covering cases a fixed regex cannot enumerate.

See more PMLE practice questions, answers explained.

Exam traps in Monitoring AI Solutions

Answers that look right on this material and are not. Each one is a distractor from a different question in the PMLE bank for this domain.

Add Model Armor prompt screening so incoming user requests are checked for injection attempts, on the basis that blocking the malicious prompt also prevents any sensitive data appearing in the reply.
Why it is wrong: Prompt screening is tempting because Model Armor does offer it, but it inspects inbound requests only; a benign-looking prompt can still trigger a tool call whose results leak into the response, so the outbound channel stays unprotected.
A safety filter detects prompt-injection and jailbreak attempts, whereas Model Armor only classifies content into harm categories such as hate speech.
Why it is wrong: Tempting because both screen content, but it reverses the roles: injection and jailbreak detection is Model Armor's territory, while harm-category scoring is what the safety filter provides.
Model Armor inspects only inbound prompts by design, so the team must add a separate regular-expression post-processor to catch sensitive values in the response.
Why it is wrong: Tempting because many gateways are inbound-only and regex post-processing is a real pattern, but it is wrong: Model Armor already screens responses, so a bolt-on regex layer is not required to cover the output path.

Examworthy is not affiliated with or endorsed by Google Cloud. Original, blueprint-aligned practice material only.