SQL injection remains a major threat to web application security, compromising data integrity and service availability. Leveraging advanced feature extraction and semantic modeling capabilities, deep learning and large language models have emerged as pivotal approaches for SQL injection detection. This paper systematically reviews recent advances in this field, with a focus on architectural designs, detection performance, and application contexts of deep neural networks and hybrid models. The review also examines datasets, evaluation metrics, and the relationship between architectural choices and detection efficacy, highlighting the benefits and deployment challenges of large language models, and concludes by identifying key issues and future research directions.