EduGuardBench: A Holistic Benchmark for Evaluating the Pedagogical Fidelity and Adversarial Safety of LLMs as Simulated Teachers

Most AI teacher models struggle with basic teaching competence: Across 14 leading models tested, "incompetence remains the dominant failure mode," meaning these systems often fail to demonstrate proper pedagogical skills even before considering safety concerns.

Bigger models aren't always safer: The research found a "counterintuitive scaling paradox" where mid-sized models were sometimes more vulnerable to adversarial attacks than larger ones, challenging the assumption that larger AI systems are automatically more trustworthy and secure.

The best AI teachers turn harmful requests into learning opportunities: Models with the highest safety ratings demonstrated an "Educational Transformation Effect"—they could convert inappropriate requests into legitimate teaching moments rather than simply refusing them, representing an advanced form of responsible AI behavior.

Academic misconduct is a unique teaching-specific risk: Unlike general AI safety benchmarks, EduGuardBench identified harms specific to educational contexts, particularly the risk that AI teachers could help students cheat or engage in academic dishonesty—a vulnerability that standard safety tests might miss.

We need specialized benchmarks for AI in professional roles: The paper demonstrates that existing safety evaluations don't adequately measure whether AI systems can actually perform professional duties competently and ethically, suggesting that different professions require tailored evaluation approaches.