The invention belongs to the technical field of software testing. The invention particularly relates to a program path sensitive grey box testing method and device. The method comprises the steps thatin the offline training stage, vulnerability mode learning is conducted on a sample data set through a deep neural network, a classifier of a program execution path is obtained, and a sample data setpackage comprises vulnerability program path sample data and vulnerability-free program path sample data; and in an online test stage, the classifier is integrated into a fuzzy test tool to guide a seed file to perform a selection test, seed input triggering a vulnerability path is preferentially selected to perform the test, endowing the test with a plurality of variation energies to execute corresponding variation times, and performing cyclic execution until interruption. The method fills up the blank of vulnerability path sample influence analysis, does not depend on a complex dynamic analysis technology, does not bring about a large overhead problem, can be effectively combined with other grey box test technologies, improves vulnerability discovery efficiency, can be directly suitablefor a binary program, does not depend on a source code, and is high in applicability.