Building Effective Safety Guardrails in AI Education Tools