Abstract: The rapid advancement of artificial intelligent generated content (AIGC) technologies and the growth of the digital economy have led to an explosion in heterogeneous content creation, posing increasingly complex challenges for content moderation (CM). While large language models (LLMs) have made significant progress in CM, most existing frameworks are limited to single-modality inputs (e.g., text), and are therefore inadequate for detecting nuanced or latent risks in multimodal data. Moreover, the computational demands of LLMs hinder their deployment on edge devices with limited resources. To address these challenges, we propose AllGuard—a lightweight, scalable multimodal CM framework optimized for edge deployment. AllGuard integrates a modular pipeline capable of handling text, image, audio, and video inputs through multimodal tokenization and prompt-based safety reasoning. We construct and annotate a high-quality multimodal dataset based on a comprehensive safety taxonomy, and fine-tune LLM using LoRA (Low-Rank Adaptation)-based adaptation for efficient model specialization. Empirical results show that AllGuard achieves state-of-the-art performance across multiple modalities, with an overall accuracy of 91.58% and strong generalization on edge hardware, maintaining over 90.66% accuracy on both the Jetson Orin and Orange Pi 5 Plus platforms. Our work provides a practical, robust, and deployable solution for real-world CM, promoting the safe and responsible use of AIGC applications.