<![CDATA[Large Language Models (LLM) deployed in sensitive public services, particularly staff selection chatbot systems, are critically vulnerable to prompt injection and jailbreak attacks, yet existing guardrail solutions remain homogeneous and prone to single points of failure. This research designs, implements, and evaluates a novel multi-layered guardrail architecture based on a defence-in-depth strategy, systematically integrated across the input, retrieval, and output stages of a Retrieval-Augmented Generation (RAG) pipeline. A comparative experimental study was conducted, benchmarking the proposed architecture against state-of-the-art baselines, Llama Guard and NeMo Guardrails, across security and performance metrics. The proposed architecture achieved a Harmful Refusal Rate of 99.5% and demonstrated confirmed robustness against prompt injection, jailbreak, and data leakage vectors as validated by the Garak Vulnerability Scanner. Despite the inherent latency trade-off of layered processing, the system maintained superior efficiency with a median latency of 8.29 seconds, outperforming Llama Guard (8.80 s) and NeMo Guardrails (8.56 s). These findings establish that a defence-in-depth guardrail approach delivers an optimal balance between security and performance, and contribute a replicable architectural framework to the field of AI security for hardening LLM-RAG systems in high-stakes, public-sector deployments.]]>
Copyrights © 2026
