A GIS mapping unit is cataloging prompt-injection types for the municipal AI risk register. What should that taxonomy emphasize for governance value?
Select an answer to reveal the explanation.
Short Explanation
Risk registers love clear buckets: direct, indirect, jailbreak, tool-exfil—each with a city example. That taxonomy beats a vague “AI weirdness” sticker or a note about chair brands. Cloud logos do not replace the list.
Full Explanation
A governance-ready prompt-injection taxonomy names actionable classes—direct, indirect, jailbreak/persona-override, and tool-exfiltration patterns—each with civic examples so owners can map controls and residual risk. GIS mapping units adding these to the municipal AI risk register create shared language for audits, red teams, and procurement questionnaires. Clear subtypes beat vague labels when councils ask what could go wrong and what is already mitigated.
Cataloging only the brand of office chairs near plotters fails by concept: furniture has no bearing on injection risk taxonomy. Claiming taxonomy is unnecessary if any cloud vendor is used fails because vendor choice does not remove the need to classify and control AI input threats. A single undefined “AI weirdness” label with no subtypes fails because it cannot drive specific controls or metrics.
Exam caveat: Taxonomies should stay practical—too many micro-classes become shelfware—yet must be specific enough to assign different controls (for example, document sanitization for indirect injection versus tool allowlists for exfiltration). Update examples as new municipal assistants launch. Exam answers that dismiss classification or substitute unrelated facilities trivia are distractors.
Operational check: Publish a one-page register excerpt listing the four named classes with one city example each, assign a control owner per class, and review quarterly whether incidents map cleanly into those buckets.