Operationalizing Child-Language Jailbreak Prompts: A Methodological Framework for Controlled Transformation and Annotation
Abstract
Large Language Models (LLMs) are increasingly accessible to children through publicly available conversational AI systems, creating new challenges for AI safety and content moderation. While existing jailbreak research has primarily focused on adult-oriented adversarial prompting, considerably less attention has been paid to child-language interaction patterns and their potential influence on moderation robustness. This paper proposes a methodological framework for the controlled transformation and annotation of child-language jailbreak prompts. Building on prior research in jailbreak taxonomy, prompt transformation, child-safe AI, and pragmatic communication theory, the framework defines six child-language-oriented transformation categories: Roleplay, Non-standard linguistic variation, Indirect-request, Meta-prompt, Mentor-mode, and Emotional framing. The proposed workflow transforms adult-oriented benchmark prompts into child-language variants while preserving their underlying objective through semantic anchors and intent-preservation validation. Designed for use with established jailbreak benchmarks, including HarmBench, the framework emphasizes reproducible prompt transformation, annotation consistency, and benchmark adaptation rather than classifier development. The primary contribution is a structured methodology that supports future child-centered AI safety evaluation, moderation research, and benchmark extension workflows.