ADVERSARIAL PROMPT INJECTION DETECTION SYSTEM

This paper presented the design and implementation of an Adversarial Prompt Injection Detection System, a security solution developed to protect Large Language Models (LLMs) from malicious prompt injection attacks. The proposed system integrates rule-based pattern matching with Google Gemini AI to identify and classify potentially harmful prompts before they are processed by the language model. The implementation results demonstrate that the hybrid detection approach effectively identifies a wide range of prompt injection attacks, including instruction override, jailbreak attempts, role manipulation, prompt leakage, and hidden command injection. By combining regular expression-based detection with semantic analysis, the system provides improved detection accuracy while maintaining efficient real-time performance. The proposed solution demonstrates the practical application of artificial intelligence in enhancing the security and reliability of LLM-based applications. The lightweight Streamlit interface and modular architecture make the system easy to deploy, maintain, and extend. Overall, the developed framework provides an effective security layer that helps reduce the risk of adversarial prompt manipulation and supports the safe adoption of AI-powered systems across various application domains.

[1] S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2021. [2] T. Brown et al., "Language Models are Few-Shot Learners," Advances in Neural Information Processing Systems (NeurIPS), vol. 33, pp. 1877–1901, 2020. [3] J. Wei et al., "Finetuned Language Models Are Zero-Shot Learners," International Conference on Learning Representations (ICLR), 2022. [4] Google DeepMind, "Gemini API Documentation," Google AI for Developers, 2025. Available: https://ai.google.dev [5] Streamlit Inc., "Streamlit Documentation," 2025. Available: https://docs.streamlit.io [6] Python Software Foundation, "Python 3.11 Documentation," 2025. Available: https://docs.python.org/3 [7] OWASP Foundation, "OWASP Top 10 for Large Language Model Applications," 2025. Available: https://owasp.org/www-project-top-10-for-large-language-model-applications/ [8] N. Carlini, F. Tramer, E. Wallace, M. Jagielski, A. Herbert-Voss, K. Lee, A. Roberts, T. Brown, D. Song, U. Erlingsson, A. Oprea, and C. Raffel, "Extracting Training Data from Large Language Models," USENIX Security Symposium, pp. 2633–2650, 2021. [9] Hugging Face Inc., "Transformers Documentation," 2025. Available: https://huggingface.co/docs/transformers [10] A. Greshake, M. Abdelnabi, S. Mishra, and others, "More than You've Asked For: A Comprehensive Analysis of Novel Prompt Injection Threats to Application-Integrated Large Language Models," arXiv:2302.12173, 2023.