A science-center laser exhibit kiosk user tries role-play tricks to make the model give unsafe laser-pointer instructions it was told to refuse. What is that attempt called?

Jailbreaking is the deliberate attempt to bypass prompt-only refusals—often via role-play, hypothetical framing, or “ignore previous instructions” tricks—so the science-center laser-exhibit kiosk would emit unsafe laser-pointer instructions it was told to refuse. The concept matters because refusal text sitting only in a system prompt is soft policy: public kiosk users will probe for ways around wording when no stronger application control exists. Inventing a new AWS service name does not describe the attack and adds no protective control for unsafe laser guidance. Fine-tuning updates weights and is not what a walk-up kiosk user does with role-play at the exhibit floor. Prompt Management stores and versions prompts for operations; a jailbreak attempt is an adversarial interaction, not a stored prompt version. Exam caveat: prompt-only safety is necessary but insufficient for public kiosks—add application guardrails, topic denials, and monitoring rather than relying on polite system text alone. Operational check: replay known jailbreak patterns against the kiosk prompt and verify unsafe laser instructions remain blocked at the application safety layer before public hours. Capture attempt logs so educators can tighten denied topics after each probe wave without waiting for a real injury-adjacent answer to slip through.