Educational safety performance varies widely across LLMs, with GPT-5.6-Luna the safest and Qwen3-8B among the weakest under educational adversarial probing¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? GPT-5.6-Luna achieves the lowest overall ASR (11.4%), followed by Gemma-4-12B-it (18.3%) and Gemini-3.5-Flash (29.8%), while Qwen3-8B (63.1%) and Ministral-3-8B-Instruct (62.0%) show the highest ASRs. → Park 2026

Evidence¶

Park 2026¶

Park, J., Han, J., Yoo, H., Ahn, S.-Y., Yoon, J., & Oh, A. (2026). EduZone: A Framework for Evaluating LLM Safety for K-12 Students and Teachers. arXiv preprint. https://arxiv.org/abs/2608.02024 doi:10.48550/arxiv.2608.02024

q2 · i? · design · r2

Model comparison in the EduZone evaluation of four proprietary and six open LLMs. Qwen3-8B and Ministral-3-8B-Instruct exhibit the highest ASRs, indicating the weakest educational safety performance among evaluated models.

"GPT-5.6-Luna achieves the lowest ASR (11.4%), followed by Gemma-4-12B-it (18.3%) and Gemini-3.5-Flash (29.8%)."

Discussion¶

Related Claims¶

• Cross-model validation shows consistent degradation under multi-strategy attack across Llama-3.3-70B, GPT-4o-mini, and Claude-3.5-Haiku, with significant differences between models — related
• Boundary compliance (S3) drives nearly all inter-model safety divergence, with a 53-point gap between the safest and least safe models — related
• FATE benchmarks rank Gemini 2.5 Flash highest among commercial AI tutors, with actionability weakest for all models — related