The autonomous pentest system red teams keep sharing: a community-maintained fork that lays the attack surface bare
What it does
RuoJi6/ARTEX is an independently maintained version of the open-source autonomous penetration-testing system ARTEX, with a Go back end and a Next.js front end. In plain terms, it turns "let an AI agent autonomously explore a target network, find weaknesses and map asset relationships" into a platform with a real interface. It preserves the original Autumn-27/ARTEX commit history and the AGPL-3.0 license; its release packages come from this repository, and the Docker image is built by this repository's own GitHub Actions. For anyone wanting to stand up their own attack-surface exploration environment, it is an out-of-the-box choice.
Why it is climbing fast
The original ARTEX is already well known in autonomous pentest circles, but its upstream update pace is uneven, so several community maintenance forks appeared. This RuoJi6 version, created in October 2026, quickly reached 293 stars. It is gaining traction because it offers a "I can self-host and stay in control" guarantee: release packages, SHA256SUMS and the Docker image all flow through this repo's CI automatically, so the trust chain is clear. For defenders, the value of such a tool is not "go attack someone" but using it as a mirror: feed your own network in, watch which path the AI walks in from, and close the blind spots.
Technical highlights
Its interface breaks the whole pentest process into fine detail. The dashboard shows an overview, token consumption and an activity stream. The task list and execution view expose every tool call inside each session. The discovery chain and the asset coverage graph use a force-directed layout to make "which assets got touched and how far coverage reached" instantly readable, with tested nodes highlighted and collapsible nodes. It can sync asset data directly from ScopeSentry to avoid duplicate collection, and intercept approvals plus human-in-the-loop chat ensure critical actions need a human nod. Back-end logs and LLM configuration keep the whole run auditable and controllable. In short, it balances "AI explores autonomously" with "a human can stop it anytime and review it afterward" rather well.
Who it is for
The boundary must be clear: use it only on targets you own or have written authorization for. For red teams, it is a base for authorized exercises and fast attack-surface exploration. For blue teams and defenders, it is the best "attacker-view" teaching tool: put your own assets in, watch which entry points the AI prioritizes, which is more intuitive than ten articles. For pentest-tooling developers, its front-end components, approval flow and asset graph are ready-made reference implementations.
Quick start
Download the release package from the repository, or run the officially built Docker image ruoji6/artex. If you want the interface without running real scans, the project also ships an online demo that runs only the front end with mock data (built with NEXT_PUBLIC_MOCK=1), no database and no model calls.
A practical example workflow
Say your team runs an internal web app and wants to know how bad an autonomous agent could be. You stand up ARTEX in a sandbox that mirrors your staging network, scope it to the authorized hosts, and let it explore. The dashboard fills with discovered assets, the coverage graph shows which subdomains the agent reached, and the session view reveals the exact tool calls it used to get there. Your blue team reads that path and finds three forgotten endpoints that were indeed exposed—exactly the blind spots you feared. Because every critical action went through intercept approval, nothing outside the sandbox was touched, and the back-end logs give you a clean write-up for the next stand-up. That is the loop defenders actually want: safe, observable, repeatable.
Where it still needs you
Keep the guardrails on. This is a capable autonomous tool, so run it only in scoped sandboxes with written authorization, never against production you do not own. The force-directed graph is great for a feel but should be exported and tracked over time, not eyeballed once. And remember the AI's exploration is only as good as the LLM and tools you configure; a misconfigured model can miss real issues or suggest noisy ones. Treat its output as leads for humans to confirm, not as a pass or fail grade. With that discipline, it is a strong addition to a security program.
How it compares
Against the original Autumn-27/ARTEX, this fork's difference is "supply-chain verifiable": artifacts come from this repo's CI with checksums, so you are not betting on what upstream changes someday. Against commercial pentest platforms, it is open source, free, self-hostable, and can pull asset data from ScopeSentry, fitting teams on a budget that still want control. One reminder: the more capable an autonomous pentest tool is, the more strictly the authorization line must hold; any use on unauthorized targets is illegal, so confirm written permission for your targets before deploying.