Objectives: As the Korean construction industry increasingly relies on foreign workers to address labor shortages, limitations of conventional safety education, characterized by one-directional instruction and insufficient linguistic adaptability, have emerged as a critical risk factor. These approaches fail to accommodate language barriers and heterogeneous learning needs, thereby increasing accident risks and hindering workplace integration. This study proposes an intelligent safety education framework and an AI-driven training architecture for foreign construction workers in Korea, integrating perspectives from construction safety, educational technology, and artificial intelligence. Methods: The proposed framework con ceptualizes safety education as a progressive process comprising four stages: digitized, adaptive, immersive, and predictive. Based on this framework, the architecture leverages Large Language Models (LLMs) and multimodal data analysis to automatically generate multilingual, context-aware training content and to deliver personalized learning pathways according to individual proficiency levels and risk profiles. Results: An implementation case demonstrates the technical feasibility, scalability, and regulatory compatibility of the proposed system with Korean industrial safety standards. Conclusions: The findings suggest that the proposed approach can enhance safety education in multicultural construction environments and contribute to the development of a sustainable, technology-driven safety culture that supports both safety competency development and social inclusion.