A dark-sky-park reservation assistant starts emitting fragments that look like another city’s logs, not the posted park rules. What is the appropriate tester response to this possible context-manipulation leakage symptom?
Select an answer to reveal the explanation.
Short Explanation
Foreign log scraps in a park bot are smoke, not a treasure hunt. Do not poke the fire with a longer prompt to “catch it again.” Stop, leave the output alone, and call security.
Full Explanation
Dark-sky-park reservation assistants that emit fragments resembling another city’s logs are showing a possible context-manipulation leakage symptom. The appropriate tester response is to stop, not reuse the output, and notify security rather than trying to reproduce the leakage. Civic parks must not become forensic playgrounds for coaxing more foreign log text.
Lengthening the prompt to force reproduction fails by concept and by safety practice: that risks amplifying disclosure. Saving fragments as reusable fixtures fails because unexpected foreign logs are not coverage gold. Lowering temperature only and continuing to use the reply as acceptance criteria fails because parameter tweaks do not authorize treating leaked-looking content as park rules.
Exam caveat: “stop and notify” outranks curiosity. Exam items test responsible handling, not skill at extracting more training or third-party residue.
Operational check: quarantine the session, preserve minimal metadata for security, delete or lock local copies of the fragments per incident procedure, block promotion of any acceptance criteria derived from the tainted reply, and restore testing from the posted park policy only after clearance.