Vulnerability detection is an important way of improving the security of software.However,the development of the Internet makes it possible for hackers to attack software systems with new techniques.This paper focuses on a new kind of vulnerability that relates to numerical stability.It poses new threat to software security as hackers are able to bypass security protection by the new kind of vulnerability,and numerical analysis is difficult to detect such a vulnerability.In the paper,the vulnerability related to numerical stability is defined from the perspective of software behavior variation caused by numerical errors,and an automatic detection approach is proposed.The approach combines the static analysis and symbolic execution,and detects the vulnerability by three steps:symbolic extraction,static attack analysis,and dynamic attack verification with high precision values.This new approach is evaluated on a few famous open source projects.The results show that the proposed approach effectively detects the vulnerabilities related to numerical stability hidden in the real-world projects.