This proposal aims to develop a static analysis framework designed to detect and mitigate security risks in GPU programs. By leveraging program analysis and large language model (LLM) techniques tailored to GPU programs, the proposed tool will provide automated vulnerability detection, offering developers insights into potential security flaws in GPU code before deployment. Ensuring the security of GPU applications is critical as these devices are increasingly integrated into safety-critical and privacy-sensitive environments, such as cloud computing and edge devices. The proposed research will contribute to strengthening GPU program security, ultimately improving the reliability and trustworthiness of GPU-accelerated applications.