An AI Agent Workflow for Generating Contextual Cybersecurity Hints

Abstract

Large Language Models (LLMs) have been used to aide student learning across a multitude of domains such as computer science, data science, and mathematics. While these chatbots show promise, they may not perform well when quality student data is unavailable.

Hint generation for cybersecurity can be feasible with the technology we have today if we make the right simplifications. First, we need human-in-the-loop systems because the training data used to train general LLMs may not cover cybersecurity well. Second, using modular agents allows us to access dynamic data that are outside the training dataset and add specificity to the hints. In this paper, we leverage n8n, an agent deployment service, to establish the connections between our agents and Discord, the messaging system that our classroom uses to offer students a streamlined learning experience when working on interactive cybersecurity exercises. The system has undergone initial pilot testing, Hints provided were sufficient to propel students forward and make progress on their assigned work, in most cases.