US patent US12794929
Predicting security vulnerability exploitability based on natural language processing and source code analysis
Abstract
Various embodiments include systems and methods to implement a process for determining expected exploitability of security vulnerabilities. Vulnerability information corresponding to a security vulnerability is input into a multi-headed neural network. A first feature vector is output via a probability of exploitation head of the multi-headed neural network. The first feature vector is extracted from the vulnerability information and comprises a first set of features. A second feature vector is extracted from code snippets and an abstract syntax tree analyzer, with the second feature vector including a second set of features related to the security vulnerability. The two feature vectors are concatenated to produce a third feature vector, and a regression model is used to determine a probability of exploitation for the security vulnerability based at least in part on the third feature vector.