Phishing in social networks represents a significant threat to digital security, affecting both individual users and organizations. Traditional detection methods, such as blacklists and rule-based filters, are insufficient despite the dynamic and changing nature of these attacks, which requires the development of more advanced and adaptive solutions. This work aims to design a model based on Deep Learning techniques to improve phishing detection in social networks using CNN, LSTM and GRU architectures, combined with advanced contextual embeddings such as Word2Vec and BERT. The methodology was divided into five phases: collection of a representative dataset with URLs classified as phishing or non-phishing; data preprocessing using tokenization and textual cleaning; feature extraction using embeddings to transform URLs into semantic and contextual vectors; model training; and evaluation using metrics such as Accuracy, Precision, Recall and F1-Score, along with confusion matrices. The results demonstrate that the CNN model is the most effective, obtaining an accuracy of 93.42% and an F1-Score of 93.39%, in addition to presenting the lowest number of false positives and negatives compared to LSTM and GRU. In conclusion, Deep Learning techniques, particularly CNN, represent an effective solution for phishing detection in social networks, laying the groundwork for future research that considers hybrid architectures and their integration with real-time cybersecurity systems.