This paper introduces MAGENTA+, an enhanced dataset and framework for detecting Arabic Machine-Generated Text (MGT) across multiple domains. As Large Language Models (LLMs) advance, distinguishing between human-written and machine-generated text becomes crucial, especially for less-resourced languages like Arabic. MAGENTA+ expands previous work by adding a new domain (Twitter), increasing dataset size by 33%, and improving generated text quality and diversity. We conduct experiments using state-of-the-art transformer models (AraBERT) and classical machine learning techniques, addressing overfitting and model robustness. Our findings show ensemble methods effectively capture diverse aspects of text classification, achieving a 93% macro F1-score. We provide insights into domain-specific performance and challenges in detecting high-quality Arabic MGT.