c3rb3rus · security-first AI

AI agents that ask before they act.

c3rb3rus builds security-first AI that runs on local hardware. The first piece, jsaiOS, gives a model a short list of tools and a risk policy that decides what it may do, what needs a human yes, and what it can never touch.

Three shepherd dog heads outlined in orange light: the c3rb3rus mark

Projects

Small, testable pieces that grow into one platform. Each one is built to be attacked, by me first.

Pythonagent runtimein development

jsaiOS

An agent runtime where every tool has a risk level. Read-only tools run, changes need approval at the terminal, and high-risk tools are never offered to the model. Every request is audited.

How it works →
local LLMsstudy toolv1.1

cert-quiz

A certification quiz partner that runs on local models. One model writes questions, a second model checks every answer key, and review is adaptive and spaced.

How it works →
homelabwriteups coming

The lab

Full-disk encryption with remote unlock, custom Secure Boot keys, hardware security keys, and local models served LAN-only behind API keys. Writeups will cover what broke and how it was fixed.

Writeups →

Principles

Least privilege

A model gets the narrowest tool that does the job, never a shell.

Model output is untrusted

Tool calls are validated against a schema before a human is even asked.

Local first

Prompts, keys and data stay on hardware I control.

Audit everything

Every decision is logged. Argument values never are.