penetration-llm

A local-first penetration-testing agent that drives an offensive-security model entirely on-device. No cloud, no API keys, nothing leaves the machine.

What it is

A terminal assistant that runs an offensive-security model (a fine-tune of a 7B coder model) through a project-local Ollama on a loopback port. It proposes shell commands, runs them under one of three levels of human control, stores the session in SQLite, and exports a Markdown report. One loaded model plays two roles, planner and reporter, so there is no second resident model to pay for in memory.

Why I built it

I wanted an AI pentest helper I could point at my own infrastructure without sending recon data, credentials, or command output to a hosted model. Keeping the whole loop on-device made that safe, and forced a clean design: strict parsing of the model's command proposals, with a fallback for when a small model just drops a command into a code fence instead of structured output.