A Layered Security Framework Against Prompt Injection in RAG-Based Chatbots

A new three-layer security framework has been proposed to mitigate prompt injection vulnerabilities in retrieval-augmented generation (RAG) chatbots, addressing a critical issue identified by the OWASP Top 10 for LLM Applications. The framework includes a rule-based input screening layer, a provenance-based instruction hierarchy for context assembly, and an output audit layer that collectively reduce the Attack Success Rate (ASR) from 71.4% to 11.3% across models such as GPT-4o, Llama 3, and Mistral 7B, while maintaining a low false positive rate of 4.8% and a median latency overhead of 61.2 ms. This model-agnostic approach can be deployed as middleware, enhancing the security of LLM applications without requiring modifications to the underlying models, making it significant for practitioners concerned with LLM security.