Signal EduZone: A Framework for Evaluating LLM Safety for K-12 Students and Teachers

Summary

Junyeong Park, Jieun Han, Haneul Yoo, So-Yeon Ahn, Jinsung Yoon, and Alice Oh published a paper on arXiv (cs.CY) on August 3, 2026, presenting EduZone, a framework for evaluating the safety of large language models used in K-12 education by students and teachers. The authors note that while LLMs are increasingly used across K-12 education tasks, existing safety evaluations rarely examine how harmful or inappropriate content actually appears in interactions between LLMs and students or teachers. EduZone systematically combines student- and teacher-facing usage contexts, fine-grained curriculum concepts, and 6 risk categories spanning 28 subcategories covering both conventional and education-specific harms to generate contextually grounded adversarial interactions across single-turn, static multi-turn, and dynamic multi-turn settings. Using these interactions, the researchers evaluated 10 LLMs across four safety levels: refusal, safe assistance, risky assistance with safety guidance, and fully risky assistance. The results show greater vulnerability to education-specific risks and dynamic multi-turn interactions, with existing safety guardrails failing to adequately address these risks. The authors present EduZone as an automated, scalable evaluation framework to support development of safer LLMs for K-12 education.