In order to address the emerging security challenges of the industrial internet, the native adaptation advantages of semantic communication and artificial-intelligence large models were used to investigate a bidirectional empowerment scheme. The results show that the fused framework is feasible and efficient, as verified by the M/M/N queuing model, and that key performance indicators are improved. It is concluded that the proposed security large-model architecture offers a practical path for the evolution of industrial internet protection systems, and recommendations for implementation are provided.