How we train AI to uncover malicious JavaScript intent and make web surfing safer

Introduced a new AI model for Page Shield that detects specific malicious JavaScript intents (Magecart, crypto mining, malware) by parsing JavaScript into concrete syntax trees (CSTs) using tree-sitter, converting them to an AST-like structure, and feeding this into a message-passing graph convolutional network (MPGCN). The model handles obfuscated code by treating it as is and leveraging graph structure information. The TF-GNN library is used for implementation, with consideration for switching to PyTorch Geometric due to sparse matrix support limitations.