With the development of computer technology, nowadays the scale of software becomes larger and the vulnerability becomes harder to be detect. Manual audit requires a large amount of time and labor, which makes it infeasible in largescale software. Machinebased automatical vulnerability detector design therefore becomes a heated research area. We study the stateofart scalable fuzzer named American Fuzzy Loop (AFL), and aiming at its defects, such as difficulty on passing magic number tests, we analyze the reasons and give a practical solution. The improved method achieves good results in our tests.