Ensuring AI safety is paramount. Imagine training a highly effective AI safety classifier with the efficiency of a tiny model. New research demonstrates exactly this, introducing Layer Enhanced Classification (LEC), a technique that leverages the hidden states of pruned Large Language Models (LLMs) to create incredibly efficient and powerful safety filters. LEC trains a simple logistic regression classifier on the optimal intermediate layer of an LLM, outperforming even industry giants like GPT-4o on tasks like content safety and prompt injection detection. Surprisingly, smaller general-purpose models, when pruned and combined with LEC, demonstrate exceptional performance with minimal training data. This suggests an inherent ability within LLMs to extract robust features, opening doors for highly efficient, real-time safety monitoring during text generation. This breakthrough could revolutionize AI safety, making robust content filtering accessible for a wider range of applications with limited resources, ultimately enhancing trust and security in AI systems. While further research is needed to explore broader classification domains and fine-tuning possibilities, this lightweight approach holds immense promise for a safer and more responsible AI future.