Context aware inspection of generative artificial intelligence traffic using a sliding window

Abstract

A network security system (NSS) is described that performs context aware GenAI traffic inspection using sliding windows. The NSS receives an interaction between an endpoint and a GenAI model, where data of the interaction arrives at the NSS in serial order. The NSS analyzes the interaction by storing the data in chunks in serial order as the data arrives, selecting a batch of the chunks including a target chunk that fall within a sliding window, evaluating the batch of chunks with a detection module that includes a language processing model such that the chunks in the batch preceding the target chunk provide context awareness for the language processing model, and repeatedly advancing the sliding window from the target chunk to the next chunk in serial order to select the next batch. The NSS applies a security policy to the interaction based on results of analyzing the interaction.