A Context-Aware Framework for Detecting Prompt Injection in Retrieval-Augmented AI Systems

PromptShield proposes a context-aware approach to prompt-injection screening in retrieval-augmented AI systems. Its main idea is to preserve the distinction between instructions that govern the application and text retrieved as evidence. Provenance, structural analysis, semantic classification, risk scoring, and policy-aware sanitization work together to reduce the chance that untrusted passages redirect the model. The framework also requires independent tool authorization, careful logging, source-separated evaluation, and explicit measurement of benign-evidence retention. The approach is modular and suitable for a final-year CSE prototype that begins with local documents and gradually adds classifier-based screening. Because the framework has not yet been experimentally validated, its effectiveness remains a research question rather than a proven result. A completed implementation should use a well-labelled benchmark, report actual metrics, analyse errors, and state its limitations clearly. Secure RAG requires layered controls and continuous evaluation; prompt screening is valuable, but it cannot replace application-level security

[1]     P. Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks,” Advances in Neural Information Processing Systems, vol. 33, pp. 9459–9474, 2020. [2]     K. Greshake et al., “Not What You’ve Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection,” in Proc. AISec Workshop, 2023. [3]     OWASP Foundation, “LLM01: Prompt Injection,” OWASP Top 10 for Large Language Model Applications, 2025. [4]     T. Brown et al., “Language Models are Few-Shot Learners,” Advances in Neural Information Processing Systems, vol. 33, pp. 1877–1901, 2020. [5]     S. Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models,” in Proc. International Conference on Learning Representations, 2023. [6]     L. Ouyang et al., “Training Language Models to Follow Instructions with Human Feedback,” Advances in Neural Information Processing Systems, vol. 35, 2022. [7]     A. Vaswani et al., “Attention Is All You Need,” Advances in Neural Information Processing Systems, vol. 30, 2017. [8]     T. Wolf et al., “Transformers: State-of-the-Art Natural Language Processing,” in EMNLP: System Demonstrations, pp. 38–45, 2020. [9]     M. Mitchell et al., “Model Cards for Model Reporting,” in Proc. Conference on Fairness, Accountability, and Transparency, pp. 220–229, 2019. [10]  T. Gebru et al., “Datasheets for Datasets,” Communications of the ACM, vol. 64, no. 12, pp. 86–92, 2021.