We showcase a critical security vulnerability in Machine Learning (ML) by introducing a novel synonym-based attack algorithm for generating adversarial examples from text data. While ML models are widely used in applications such as chatbots and spam detectors, there is limited research on their susceptibility to adversarial attacks, especially in text classification. Our algorithm creates adversarial examples by substituting words with synonyms while maintaining semantic integrity. We evaluate the algorithm’s effectiveness on three widely used text classification models—deep neural network, convolutional neural network, and long short-term memory—using the IMDB movie review and Reuters news datasets. The impact of the adversarial examples on model accuracy is assessed, with a particular focus on the threshold parameter M, which dictates the extent of text modifications. The results demonstrate that our algorithm generates adversarial examples that remain undetectable by human reviewers and spell checkers. When tested with a threshold M of 12, model accuracy declines significantly, ranging from 38.43 to 44.16%. Accuracy deterioration intensifies with increasing threshold values. Our research highlights a significant security gap in text-based ML systems and emphasizes the urgent need for robust defenses against proposed adversarial attack.