What Is ARTEX? AI Autonomous Pentest Agent Explained: Timeline & Deploy Guide

Quick Answer

An AI autonomous pentest agent orchestration system (Go + Next.js) that needs an external LLM API to work; the original went closed-source, and mhtsec/ARTEX is the community backup of the last open release.

ARTEX is an AI autonomous penetration-testing system: not a language model itself, but an agent orchestration layer. Plug in a Claude, ChatGPT or DeepSeek API key and it runs the full pentest loop on its own — reconnaissance, vulnerability identification, attack-path planning, tool execution and result verification. It won Baidu Security Response Center's agent+ challenge in September 2026; in early October, a CrowdStrike report linked it to intrusions at South Korean banks, and the author took it closed-source. The mhtsec/ARTEX repo on GitHub is a community backup of the last open-source release: 3,757 stars and 6,127 forks in 5 days.

What Is ARTEX? The Short Answer

ARTEX is an AI autonomous penetration-testing system built by security engineer Autumn-27: a Go backend, Next.js frontend and PostgreSQL database. Its selling point is autonomy. Traditional pentest tools (e.g. Burp Suite) need a human operator at every step. ARTEX splits the job across cooperating agents: a Planner decomposes the target into discrete attack intents, parallel Workers call tools to probe, and a Main Agent handles human interaction. Findings persist in two graphs — an asset graph for discovered infrastructure and an exploration graph for goals, intents, facts and conclusions — so long engagements survive. Two caveats up front. First, ARTEX ships no language model; it is an orchestration layer. Reasoning comes from the model you plug in — Anthropic Claude in the official docs, with OpenAI ChatGPT and DeepSeek also supported (see this Claude Sonnet 4 agent guide https://yuliuswiki.com/en/claude-sonnet-4-agent for what Claude agents can do). Second, it has human-in-the-loop gates: dangerous actions pause for manual approval, which is what separates it from pure black-box attack tooling. Try the live demo first (https://artex-demo.mhtsec.com) — frontend-only mock data, no deployment needed.

Why It Exploded in a Week: Full Timeline

September 2026: ARTEX wins the agent+ offensive/defensive challenge held by Baidu Security Response Center, gaining fame in China's security community. Its repo bio still reads "Baidu agent+ challenge champion project". Late September – early October: several South Korean banks disclose intrusions. Shinhan Bank: ~25,000 customers' personal data compromised; KB Kookmin Bank: 119 records. At least nine banks were disclosed or reported by local media as targets; Korean police opened an investigation. Around October 7: CrowdStrike publishes an attribution report with "moderate confidence" that a suspected China-based actor used ARTEX together with Anthropic Claude against South Korean financial institutions; Korea's Financial Security Institute (FSI) confirmed the link via attack IPs and server logs (https://www.devdiscourse.com/article/international/3988670-crowdstrike-says-china-based-suspect-used-ai-tools-in-south-korean-bank-hacks). October 8–9: author Autumn-27 posts on GitHub: given the misuse of the tool, the ARTEX project will no longer be updated and will go closed-source, with no further public releases or maintenance. The original repo is taken down (verified by Reuters). October 8: community backup mhtsec/ARTEX is created, its README stating it is "a pure-source backup of ARTEX's last version" (https://github.com/mhtsec/ARTEX). In the next 5 days it collects 3,757 stars and 6,127 forks — forks overtaking stars, the classic signature of doomsday-backup virality: people came to preserve the code before it vanished, not to star it.

Architecture: How It Works

Technically, ARTEX is a well-built multi-agent engineering project, not magic: 1. Agent separation of concerns. The Planner plans but never touches tools; Workers execute but never decide; the Main Agent only talks to humans. This keeps long-horizon tasks on track. 2. Dual-graph memory. The asset graph stores discovered subdomains, IPs, ports and sites; the exploration graph stores each round's goals, intents, facts and conclusions. State lives in PostgreSQL, independent of any single model context window, so interrupted tasks resume. 3. Human-in-the-loop approval. Global approval logs, in-task interception approvals and in-chat approval cards — risky actions need a human click. The author drew the line early: personal learning, code research and local verification only; no real-world testing against online systems. 4. Asset sync. One-click import from ScopeSentry (domains, subdomains, IPs, ports, sites) skips repetitive recon. 5. Management console. Listens on TCP 8787 by default, with a dashboard for task overview, token usage and activity streams. In one line: ARTEX's innovation isn't any single technique — it's making the full "recon → plan → execute → verify → review" pentest loop actually run as agent engineering.

Deploy Locally in 4 Steps

Based on the backup repo README; repo verified accessible on 2026-10-13. Step 1: clone. git clone https://github.com/mhtsec/ARTEX.git && cd ARTEX Step 2: run the installer. ./install.sh — it auto-detects/installs Docker, then offers two modes: ① all-in-Docker (just set a Postgres password); ② local build (Go compiles to a single binary; use an existing DB or spin one up with Docker). Step 3: configure an LLM. ARTEX ships no model — set ANTHROPIC_API_KEY or OPENAI_API_KEY in env or the web UI (DeepSeek's OpenAI-compatible endpoint also works; see https://yuliuswiki.com/en/claude-opus-5-5-agent for Claude's agent capabilities). API costs go to the model provider, not ARTEX. Step 4: open the console at http://localhost:8787, visit /setup on first run to set the admin password, then create your first test task. Lazy option: the online demo (https://artex-demo.mhtsec.com) is frontend-only mock data. And again: only use it against targets you have written authorization for.

Where It Fits — and Where It Doesn't

Good fits: authorized pentests and red-team exercises for enterprise security teams; contracted security-service engagements; students and researchers studying multi-agent architecture — the code itself is a solid agent-engineering case study. Off limits: any target you are not authorized to test. The README's prohibitions are blunt: no real-world testing against online systems or websites. Unauthorized intrusion is illegal in most jurisdictions — open-source tooling does not legalize the act, and the author explicitly opposed illegal use and disclaimed responsibility in the takedown notice. Disclaimer: this site is not affiliated with ARTEX's original author or mhtsec; this article is a technical introduction only. ARTEX trademarks and code copyright belong to their respective owners (backup repo under AGPL-3.0). Any penetration testing requires prior written authorization from the target's owner, or you may face legal liability.

Sources

Key facts verified as of 2026-10-13: GitHub backup repo (dates, stars, license, README): https://github.com/mhtsec/ARTEX CrowdStrike attribution reporting (bank incidents, moderate confidence): https://www.devdiscourse.com/article/international/3988670-crowdstrike-says-china-based-suspect-used-ai-tools-in-south-korean-bank-hacks ARTEX shutdown timeline: https://agentbearcorps.com/article/artex-goes-closed-source-after-korean-bank-hacks-2026 Technical architecture analysis: https://cyberwarrior76.substack.com/p/artex-ai-south-korea-financial-sector Takedown announcement reporting: https://mitsloanindia.com/article/artex-developer-closes-ai-hacking-tool-after-bank-attacks/

ARTEX vs Other Pentest Tools

Dimension	ARTEX	PentestGPT	Burp Suite	Metasploit
Positioning	Autonomous AI pentest agent (full loop)	LLM-assisted pentest chat (human-guided)	Manual web vulnerability testing proxy	Exploitation & post-exploitation framework
Autonomy	★★★★★	★★★☆☆	★★☆☆☆	★★☆☆☆
Open source	AGPL-3.0 (community backup)	Open source	Commercial (free Community ed.)	Open source
Chinese support	★★★★☆ (Chinese project)	★★★☆☆	★★★☆☆	★★☆☆☆
Learning curve	Medium (Docker + API key)	Low	Medium-high	High
Cost	Free (pay your own LLM API)	Free (pay your own LLM API)	Community free, Pro paid	Free

mhtsec/ARTEX backup repo in 5 days: forks overtook stars

Stars3757

★ Forks6127

数据来源：GitHub API, checked 2026-10-13. Forks exceeding stars is unusual — most people rushed to back up the code after the takedown, not to star it for daily use.

Frequently Asked Questions

Is ARTEX an AI language model?

No. It is an agent orchestration layer (scaffolding); reasoning comes from the Claude, GPT or DeepSeek model you connect. Think of it as giving a language model hands and feet for pentesting: the model thinks, ARTEX plans, calls tools and executes.

The original was taken down — can I still get ARTEX?

Yes. mhtsec/ARTEX is a community backup of the last open-source release (AGPL-3.0), verified accessible on GitHub on 2026-10-13. But the author has stopped maintaining it — no new versions, and the original Docker image is gone, so build from source.

Is pentesting with ARTEX legal?

Only within explicitly authorized scope, e.g. your own systems or contracted targets. Using it against unauthorized systems is illegal in most jurisdictions. The README's prohibitions also ban real-world testing against online systems.

What do I need to deploy ARTEX?

A machine that runs Docker (the installer handles dependencies and PostgreSQL) plus an LLM API key (Anthropic, OpenAI or DeepSeek). Then open http://localhost:8787 and visit /setup on first run to set the admin password.

How is it different from Burp Suite?

Burp Suite is a manual/semi-automated proxy for capturing and verifying web vulnerabilities — the human operates. ARTEX is an autonomous agent that plans attack paths, calls tools and verifies results itself, with humans approving only high-risk steps. They complement rather than replace each other.

Why did the author suddenly go closed-source?

In early October 2026, after CrowdStrike's report linked ARTEX to the South Korean bank intrusions, author Autumn-27 posted on GitHub that given the misuse of the tool, the project would no longer be updated and would go closed-source. The original repo was then taken down; what circulates now is the community backup.