A municipal security operations center wants early warning when resident-facing bots keep receiving the same prompt-injection patterns. Which operational control best surfaces that threat signal?
Select an answer to reveal the explanation.
Short Explanation
Think of the SOC like a firehouse that needs the smoke alarm, not a yearly walkthrough of every hallway. When the same injection patterns keep tripping safety classifiers, those alerts should light up monitoring so responders can act before a bad day becomes a worse one.
Full Explanation
Routing repeated safety-classifier injection detections into municipal SOC monitoring and alerting is the operational control that turns per-request GenAI filters into early-warning telemetry. Resident-facing 311 and service bots will inevitably see adversarial prompt-injection attempts; when the same patterns recur, those classifier hits should feed the same dashboards, paging, and incident workflows the city already uses for network and identity threats. That linkage lets analysts correlate volume spikes, source channels, and time windows before a successful jailbreak or policy bypass becomes a resident-visible incident.
Disabling safety classifiers so dashboards stay quiet during peak hours removes the sensor entirely—silence is not safety, and peak 311 load is exactly when attackers may blend in. Relying only on monthly manual reviews of random transcripts fails the early-warning requirement: sparse sampling cannot surface bursty, repeated injection campaigns in near real time. Publishing raw injection strings on the open-data portal crowdsources the attack recipe, helps adversaries refine payloads, and creates unnecessary public disclosure risk without improving SOC response quality.
Exam caveat: know the difference between applying a safety classifier at inference time and operationalizing its signals. AIP-C01-style items often test whether you treat GenAI security events as first-class SOC inputs rather than one-off application logs. Classifier presence alone is insufficient if detections never reach monitoring and response.
Operational check: confirm repeated injection labels emit structured events into the municipal SIEM or CloudWatch alarms, that alert thresholds cover same-pattern recurrence (not only single hits), and that on-call runbooks name ownership for resident-facing bot injection surges—including quarantine or tighter guardrail actions when thresholds fire.