Large language models (LLMs) are powerful, but they need safety measures to prevent misuse. Current approaches to filtering inappropriate content often rely on equally massive AI models, creating latency and high costs. However, new research suggests a leaner approach can be just as effective. Imagine a security guard who can screen everyone entering a building quickly and efficiently. That's the idea behind using a fine-tuned Sentence-BERT model as a lightweight safety guardrail for LLMs. This research explored using a smaller, faster model based on BERT embeddings to identify unsafe prompts. The results are promising: this lightweight model achieved accuracy comparable to much larger, resource-intensive models like LlamaGuard, but with significantly lower latency. While LlamaGuard took over two minutes to process a prompt on a single GPU, the Sentence-BERT model clocked in at around 0.05 seconds. This dramatic speed improvement opens doors for deploying effective safety measures in cost-sensitive environments like classrooms or small businesses. By clustering safe and unsafe embeddings and training a classifier on these vectors, the model efficiently distinguishes between appropriate and inappropriate content. This approach, while simple, demonstrates the potential of lightweight architectures for providing robust safety mechanisms without sacrificing performance. While the initial focus was on English text inputs, future research aims to expand the model’s capabilities to other languages and modalities like speech and video. The goal is to create a highly customizable and efficient safety net for LLMs, paving the way for broader and safer AI deployment across various applications. This research suggests that we might not need to fight fire with fire when it comes to LLM safety. Sometimes, a smaller, more agile solution can be just as effective, and significantly more efficient.