opensourceprojects.dev

A broadsheet for software that doesn't ask for your email

Self-hosted penetration testing where every finding keeps its receipts
GitHub RepoViews35

Project Description

View on GitHub

Penetration Testing That Actually Shows You the Evidence

You've probably seen the demo: an AI agent confidently reports a critical vulnerability, and then someone asks for proof and the whole thing falls apart. Was it a real finding, or did the model hallucinate a juicy-looking CVE? ARTEX takes that question seriously. It's a self-hosted, LLM-driven penetration testing platform where every finding stays bound to the assets and HTTP traffic that back it up.

What It Does

ARTEX is a multi-agent offensive security tool that runs on your own infrastructure. It isn't a chat wrapper around a scanner. A planner takes a goal and breaks it into intents, workers execute those intents with real tools, and the results get written back into two linked graphs: one for what you know about the target, and one for how the test is progressing. That goal decomposition, intent dispatch, worker execution, and write-back form a closed loop that keeps advancing until the goal is proven.

The architecture is deliberately auditable. The shared asset graph tracks what exists, the per-task exploration graph tracks what was tried, what it produced, and why—and anchors link the two together. A recording MITM proxy and a browser extension capture real traffic, so findings, evidence quotes, and retests all tie back to the exact request and response. Under the hood it's a single Go binary with a Next.js UI embedded, backed by PostgreSQL, and it works with your own Anthropic or OpenAI-compatible endpoint or a local Ollama instance.

Why It's Cool

  • Findings come with receipts. This is the whole pitch, and it's a good one. A recording MITM proxy and browser extension mean the evidence isn't a summary the model wrote—it's the actual traffic. When you need to justify a finding to a client or a teammate, you have the request and response sitting right there.

  • Two graphs, one audit trail. Splitting "what exists" from "what was tried" is a smart design choice. You can see coverage of the target separately from the reasoning trail of the test, which makes it much easier to figure out where an engagement went sideways or what's still unexplored.

  • Autonomy with actual brakes. The README is refreshingly honest about the risks here. Human-in-the-loop approvals, argv-prefix rules, scope and deny gates, token and turn budgets, a target noise governor, and prompt-injection framing on tool output—that's a lot of guardrails, and it suggests the authors thought hard about what happens when you point an autonomous agent at a live target.

  • Self-hosted and model-agnostic. One binary, your own database, your choice of model provider. You're not shipping your engagement data to someone else's cloud or locking yourself into a single vendor.

The upstream project is Autumn-27/ARTEX under AGPL-3.0, and this fork keeps that lineage while packaging it for self-hosting.

How to Try It

Getting started is a few commands. You'll need Docker and a PostgreSQL password set in your environment:

git clone https://github.com/BaaRai/ARTEX.git
cd ARTEX
cp .env.example .env            # set POSTGRES_PASSWORD (and optionally an LLM key)
docker compose up -d --build    # builds from this checkout, starts PostgreSQL

Then open http://localhost:8787. On your first visit, /setup will ask for a one-time setup token printed at startup—grab it with:

docker compose logs artex | grep "Setup token"

One thing to know upfront: there's no prebuilt Docker image to pull. The image is always built from your checkout, so the first up will take a bit longer than you might expect. If you'd rather not use Docker, the README points to other installation paths—script, binaries, source, and release archives—so check the repo for those details.

Final Thoughts

ARTEX is aimed at people doing real offensive security work who want automation without giving up the audit trail—pentesters, red teams, and anyone who's been burned by a tool that reports findings it can't defend. The two-graph model and evidence-bound findings are the kind of design decisions that matter when the output has to hold up to scrutiny. It's not a magic button, and the emphasis on approval gates and budgets makes clear the authors don't think it should be. If you've been waiting for autonomous pentesting that shows its work, this is worth a look.


Follow @githubprojects for more developer tools and open source projects.

Back to Projects
Project ID: artexLast updated: November 16, 2026 at 08:59 AM