Driven by the "dual carbon" goals, the user side of the new power system faces multidimensional security risks due to the large-scale integration of new grid-connected entities such as distributed photovoltaic systems, electric vehicle charging stations, and energy storage equipment.Firstly, the challenges to equipment safety, grid security, and societal safety posed by the massive integration of these new entities are systematically analyzed, core bottlenecks at the technical and managerial levels, such as uneven equipment quality, difficulties in cross-modal data fusion, and insufficient safety protections, are highlighted. To address the heterogeneity, semantic differences, and correlation characteristics of multimodal data, a multimodal large model construction technology is proposed, incorporating data augmentation generation, cross-modal semantic alignment, and parameter-efficient fine-tuning methods to enhance risk perception accuracy. Furthermore, integrated security prevention and control strategies are introduced from three perspectives: grid optimization, intelligent operation and maintenance, and resilience enhancement. These measures provide theoretical support and practical pathways for the security governance of the new power system, facilitating the coordinated development of energy transition and security under the "dual carbon" objectives.