A noticeable departure from conventional methods is evident in the dynamic domain of technology, enabling users to seamlessly transfer entire code snippets into generative AI models for comprehensive error correction and modification. This fundamental change in interactions brings about crucial challenges, primarily emphasizing preserving codebase integrity and safeguarding sensitive data. This paper introduces a machine learning-powered sensitive data loss prevention firewall explicitly designed for generative AI models. A pre-trained CodeBERT model was fine-tuned using transfer learning, leveraging its word embedding and attention mechanism capabilities to perform the downstream task of code-text filtering. The model was trained on a diverse dataset comprising Java and Python code samples sourced from GitHub and textual data from the Ubuntu Dialogue Corpus on Kaggle. Thus, a novel code-text filtering system was developed, effectively separating code from text and comments within documents and then blocking this code from proceeding to generative AI applications.