A parking authority worries that crafted adversarial prompts could pull internal enforcement-policy text out of an assistant that has that corpus in context. What threat is that?
Select an answer to reveal the explanation.
Short Explanation
If the assistant can see the internal policy binder, a sneaky prompt might try to walk that binder out the front door. Limit what the model can retrieve, and watch for prompts that fish for confidential text.
Full Explanation
Crafted adversarial prompts that try to pull internal enforcement-policy text out of a parking authority assistant that has that corpus in context are an exfiltration-style AI interaction threat: sensitive policy content leaves through the dialogue channel. If the model can retrieve or has been given internal playbooks, fee schedules, or enforcement tactics, an attacker or curious user may coerce disclosure via jailbreak-like or extraction prompts. Mitigations include least-privilege retrieval, redacting confidential sections from the assistant’s corpus, output filters, and monitoring for fishing prompts.
A flooded underground garage sump pump is a facilities flooding issue, not policy exfiltration via AI. A missing handicap-sign inventory spreadsheet is a records/operations gap about physical assets, not adversarial extraction from an assistant. A burned-out LED on a street pay station is hardware maintenance, unrelated to prompt-based content leakage.
Exam caveat: this threat assumes sensitive material is available to the model; if the corpus never contains confidential policy, extraction of that policy cannot succeed—so classification and corpus hygiene matter as much as prompt defenses. Operational check: inventory what the assistant can retrieve, strip or segment confidential enforcement guidance, test extraction prompts, and alert on repeated attempts to dump policy text.