<p>Social bots are a known problem in today’s society. They are influenced by a variety of factors, ranging from the presence of bots to a lack of interaction between bots and users. This paper proposes a cross-platform approach for the detection of social bots based on profile metadata and text embeddings, applied to Twitter, Mastodon, and Bluesky user accounts. The resulting model achieves 97.39% accuracy in a four-class classification task, outperforming several established baselines, including graph-based and federated approaches while being computationally efficient. The primary contribution of this work is the demonstration that user features can support effective bot classification across heterogeneous and decentralized environments, demonstrating the feasibility of cross-domain generalization at scale. We additionally present a novel dataset that combines self-identified bot and non-bot accounts from decentralized platforms.</p>