County jail medical chatbot logs show a user message trying to force the bot into an “admin mode” that unlocks restricted tools. What prompt-injection pattern is this closest to?
Select an answer to reveal the explanation.
Short Explanation
“Flip into admin mode” is the social-engineering cousin of asking for the master key. That message is hunting elevated powers, not visiting hours. Call it a privilege-escalation style injection and keep the tools locked.
Full Explanation
Privilege-escalation style prompt injection seeks elevated assistant capabilities, restricted tools, or “admin modes” that bypass normal civic limits. A county jail medical chatbot user message that tries to force an “admin mode” unlocking restricted tools matches that pattern: the goal is not ordinary information but expanded power inside the agent. Recognizing escalation-oriented injection is critical where medical, custody, or justice tools can access sensitive records.
A normal request for published visiting hours fails as a classification because it asks for public information without seeking elevated capabilities. A benign spelling correction on a discharge summary is editorial help, not privilege escalation. A scheduled database vacuum with no user text is a maintenance job outside the prompt-injection taxonomy.
Exam caveat: “Admin mode” jailbreaks are a recognition class; real authorization must never depend on the model believing a persona switch. Tool permissions should be enforced in application code and identity systems, not solely in prompt text. Jail/medical contexts heighten confidentiality and safety stakes if escalation succeeds even partially.
Operational check: Attempt a staged “enable admin mode / unlock restricted tools” probe in a non-production jail medical bot, confirm tools remain denied without staff credentials, and alert when logs show repeated escalation phrases.