Abstract
Modern malware evolves continuously, posing persistent challenges to cybersecurity. Conventional classification approaches typically group malware by its primary objective, emphasising dominant behaviours while overlooking the complex and overlapping strategies common in real-world attacks. Here we present DECODE (DEep Classification Of Dynamic Exploits), a proportional multi-label, context-aware framework that combines object detection, explainable artificial intelligence (XAI), and agent-based large language models (LLMs) to deliver interpretable and comprehensive malware analysis. DECODE introduces the first object detection dataset specifically for malware classification, generated through an automated annotation pipeline that removes the need for manual labelling and remains effective even for visually indistinguishable malware features. To improve attribution reliability, we extend Gradient-weighted Class Activation Mapping (Grad-CAM) with a Bayesian formulation, enabling uncertainty-aware visualisation of discriminative regions linked to multiple categories. The regions identified through object detection are subsequently mapped to their corresponding API call sequences and interpreted via a multi-agent reasoning module, which incorporates critique-and-verification loops to reduce hallucinations and bias. Experimental evaluation shows multi-label and binary classification accuracies of 0.8513 and 0.9380, respectively, outperforming conventional deep learning baselines. By combining visual localisation, proportional multi-label scoring, and human-readable behavioural narratives, DECODE enables malware to be classified not only by intended impact but also by fine-grained structural and behavioural traits, offering a richer understanding of complex threats.