A city SOC must improve threat coverage for the permitting chatbot. Which action best addresses the AI threat landscape?

Improving threat coverage for a municipal permitting chatbot requires cataloging AI-specific threats such as prompt injection and data poisoning against the chatbot and its knowledge sources. Those threats target the application and corpus layers: attackers manipulate instructions, exfiltrate sensitive permit data, or corrupt retrieved guidance that staff or applicants may trust. Cataloging them works because SOC detection engineering and threat hunting need named use cases—without an AI threat catalog, coverage defaults to legacy network patterns that miss conversational abuse.

Extending prompt injection and data poisoning detection to network IDS signatures for permitting database replication traffic fails because IDS signatures inspect network packets and replication streams, not prompt semantics or retrieval content, so they cannot detect manipulated instructions or corrupted guidance. Prioritizing prompt injection and data poisoning tests for chatbot TLS ciphers during quarterly encryption compliance audits fails because TLS ciphers only provide transport encryption and have no prompt behavior to exploit or poison, so an encryption compliance audit tests the wrong layer for AI threats. Hardening chatbot outputs against prompt injection and data poisoning by rotating API keys for daily permitting inquiry sessions fails because rotating keys only refreshes authentication material and does not validate prompts, filter outputs, or detect corrupted knowledge sources.

Exam caveat: cataloging threats is necessary but not sufficient—each entry needs owners, detections or compensating reviews, and test cases. Operational check: add prompt-injection and corpus-poisoning scenarios to the SOC use-case catalog, map them to logs (prompt gateway, retrieval, auth), run tabletop or purple-team exercises against the permitting bot, and verify alerts fire before declaring coverage improved.