A method for assessing a cybersecurity risk of a software object includes generating an abstract syntax tree (AST) for a software object, and determining that the AST is insufficient to identify, to a specified confidence level, a cybersecurity risk of the software object. In response to determining that the AST is insufficient to identify the cybersecurity risk of the software object, a graph convolutional neural network (gCNN) is executed, based on the AST, to produce a set of features for the AST and to produce a probability of maliciousness of the software object based on the set of features. A signal representing an alert is sent, based on the probability of maliciousness, if the probability of maliciousness exceeds a pre-defined threshold.