The security of machine learning is of great importance. Image vulnerability has been known in the literature for a long time. This paper is concerned with the text vulnerability. The word2vec is widely used to produce the word embedding, which plays an important role in natural language processing. The quality of word embedding affects the performance of the neural network. Using word embedding is to find the useful information that may exist between the individual words. This paper proposes a method to alter the original text, whose word embeddings also change. And the adversarial samples are able to make the classifier make a mistake.