Large language models (LLMs) are impressive, but their ability to generate harmful content is a serious concern. Aligning these models with ethical guidelines typically requires extensive resources and complex methods like supervised fine-tuning (SFT) or reinforcement learning from human feedback (RLHF). But what if there was a simpler, less resource-intensive way? Researchers have developed a clever method called DAPA (Decoupled Alignment for Robust Plug-and-Play Adaptation), which enhances LLM safety without the heavy lifting of traditional alignment techniques. DAPA leverages knowledge distillation, a process where a "student" model learns from a well-aligned "teacher" model. The key insight? DAPA pinpoints the most crucial parts of the teacher's knowledge related to ethical behavior. By selectively transferring these "alignment modules," DAPA boosts the safety of unaligned LLMs. Surprisingly, modifying a small fraction of the model's parameters—as little as 3.25%—can make a substantial difference. In tests across 17 different LLMs, DAPA increased the models' ability to deflect harmful prompts by an average of 14.41%, sometimes reaching improvements as high as 51.39%. Even better, this boost in safety doesn’t come at the cost of performance. DAPA-aligned models retain their core abilities, like generating coherent text and solving complex problems, with minimal degradation. This breakthrough is especially significant for developers who lack the resources for extensive fine-tuning. DAPA offers a practical and efficient way to make LLMs safer, paving the way for wider adoption in real-world applications. While the current version of DAPA modifies a relatively small portion of model parameters, future research aims to refine the process, making it even more lightweight and efficient. This work is a critical step towards ensuring the responsible development and deployment of increasingly powerful AI models.