This preprint presents SEMA-GUARD, a semantic and graph-based framework for vulnerability detection in assembly code. The proposed approach combines control-flow graph representations with semantic features derived from low-level program behavior, including stack manipulation, memory access patterns, taint propagation, and control-flow transitions. These enriched graph representations are processed using Graph Neural Networks (GNNs) to identify vulnerable and non-vulnerable functions directly from compiled code. The framework is designed for binary analysis scenarios where source code is unavailable, such as firmware analysis, malware investigation, reverse engineering, and third-party software auditing. Experimental evaluation was conducted using a dataset derived from the Juliet Test Suite, where source programs were compiled into assembly code and transformed into function-level control-flow graphs. Results demonstrate that incorporating semantic information improves vulnerability detection performance compared to structural and statistical baselines. This repository contains the preprint manuscript describing the SEMA-GUARD framework, methodology, experimental design, results, and discussion. The accompanying implementation and dataset artifacts are released separately to support reproducibility and future research in binary vulnerability analysis, graph learning, and AI-assisted cybersecurity. Keywords: Vulnerability Detection, Binary Analysis, Assembly Code Analysis, Graph Neural Networks, Software Security, Program Analysis, Cybersecurity, Control-Flow Graphs, Semantic Feature Extraction, Machine Learning for Security. Preprint version. This manuscript has been under formal peer review.