The goal of this project is to combine whiterabbitneo, an open source model specialized in cyber security, and PentestGPT, a prompting technique for GPT 4 that allowed GPT 4 to be in the top 1% of hack the box users, to create a fully open source solution for pentesting. For this repo, my goal is to beat one box in hackthebox.
However, when working with PentestGPT with models that are not GPT4, like whiterabbitneo, the outputs tend to not always follow the structure we want. For example, when updating a todo list, the items might dissapear or it may not output a todo list in the response at all. To prevent this, I added outlines for constrained generation to bridge the gap between whiterabbitneo and gpt 4. However, currently we are not there yet