This paper proposes a novel approach for detecting vulnerabilities in source code written in C and C++, leveraging large language models (LLMs). Specifically, the study introduces a new model called RoS-Dex, based on ensemble learning techniques and comprising two main components: Code Understanding (CU) and Vulnerability Encoder (VE). Accordingly, the CU module is developed using code embedding techniques and a transformer-based architecture, enabling it to capture the semantic features of source code comprehensively, while the VE module focuses on encoding vulnerability-related features, thereby improving classification performance. In the experimental evaluation, the RoS-Dex model demonstrated effectiveness not only on a single dataset but also across four datasets with different structures and characteristics, including REVEAL, FFMQ+QEMU, BigVul, and RealVul. Furthermore, the RoS-Dex model also showcased its adaptability by successfully passing cross-data validation tests—one of the most rigorous evaluation methods that very few LLM-based models have managed to pass successfully. These results highlight the strong potential of the proposed model for real-world applications and pave the way for future research in C and C++ vulnerability detection.