Cybersecurity

Taking Down ARTEX Backfired: Cloudflare Says 50-Plus AI Hacking Frameworks Probed Its Customers in Three Days

The creator of the open-source agentic attack tool pulled it on Oct. 8. Cloudflare now counts more than 12,000 related repositories, a 30-fold jump in traffic from AI pentesting tools since Oct. 1, and real break-ins in which stolen CI/CD tokens were chained by the software.

DailyMech Staff

Taking Down ARTEX Backfired: Cloudflare Says 50-Plus AI Hacking Frameworks Probed Its Customers in Three Days
LoboStudioHamburg / Pixabay — Pixabay Content License

Rack-mounted network switches with green status lights and rows of white, red and green Ethernet cables plugged into a patch panel

Illustrative photo of network switches and patch cables in a server rack. It is not Cloudflare equipment.

Five weeks ago, the developer behind ARTEX, an open-source system that uses AI agents to run penetration tests on their own, deleted the project. "The malicious exploitation of this tool runs entirely counter to its original intent," the creator, known as Autumn-27, wrote in Chinese, according to Cloudflare.

The takedown did not work. In a report published Nov. 13, Cloudflare's threat intelligence team, Cloudforce One, said the number of AI-driven hacking frameworks has "exploded" since then. In the three days before publication alone, the company said, more than 50 distinct framework families tried to scan Cloudflare customers around the world.

"Defenders are struggling to keep up with the volumetric nature of these autonomously guided attacks," wrote the authors, Blake Darché, Myles Sloan and Jacob Crisp.

At a glance

Item Detail Source
ARTEX takedown Oct. 8, by its creator, Autumn-27 Cloudflare
Repositories since More than 12,000; several hundred unique forks; 204 public forks with their own code changes Cloudflare
Framework families seen More than 50 in three days Cloudflare
IP addresses running them Thousands, rising daily Cloudflare
Traffic growth More than 30x since Oct. 1; more than 10x since Oct. 8 Cloudflare
ARTEX share Over 60% of known AI pentest orchestrators, as of Nov. 12 Cloudflare
Real incidents Multiple, including stolen CI/CD tokens chained across cloud accounts Cloudflare

What these tools are

Penetration testing is the legitimate business of attacking a system to find its weak spots before criminals do. The tools Cloudflare describes take that work and hand it to teams of AI agents.

According to the report, frameworks like ARTEX can plug into Model Context Protocol (MCP) servers and connect standard hacking tools "in novel ways." They map a target's assets onto a graph. Then several agents work in parallel to "enumerate, verify, prototype, and attack at the speed of inference." Cloudflare's summary is blunt: "These tools turn compromises into a hill-climbable problem, which agents are specifically good at."

Cloudflare says it is tracking more than 50 such agents. It named the five most popular, each with more than 1,000 GitHub stars and each fully autonomous:

  • ARTEX and its forks, a planner with parallel agents that runs on the operator's host or in a container with network and root access.
  • LuaN1aoAgent, which uses supervisor, planner and executor agents and has built-in SSH and Chisel tunneling.
  • Decepticon, an orchestrator that sends work to specialist sub-agents from a Kali Linux container running as root, with skills for Active Directory and cloud pivoting.
  • Cairn, a dispatcher that hands tasks to existing AI coding harnesses, including Codex, Claude Code and Pi.
  • PentestGPT-Agent, a supervisor paired with one executor agent.

Cloudflare calls the shift "doctrinal." In its telling, human skill is no longer the limit. Parallel agents move faster than one operator, a single person can hit several organizations at once, and small configuration mistakes get found because the software tests exposed services "relentlessly and systematically."

A hydra effect

The most striking finding is what happened after the takedown. Cloudflare's charts show the number of unique tool instances climbing steeply around Oct. 8. The company says the removal, "and widespread reporting of it," likely caused a "hydra effect": cutting off the original produced hundreds of new forks and lookalikes.

Today, there are over 12,000 repositories tied to ARTEX, with several hundred unique forks that trace their lineage to the original, according to the report. Cloudflare also counts "hundreds" of other agentic attack frameworks created since Oct. 8. Even so, ARTEX and its descendants still made up more than 60% of the known tools as of Nov. 12.

What an attack looked like

Cloudflare says that over the last eight weeks, Cloudforce One has been called into "multiple" incident investigations involving ARTEX and its offshoots.

The pattern it describes is less about clever AI and more about old mistakes moving faster. In several cases, an attacker first got into a customer's account at a major cloud provider, then pivoted straight into that customer's Cloudflare environment. The common root cause, Cloudflare says, was "static, non-expiring tokens" sitting inside CI/CD pipelines, the automated systems that build and ship software, with broad access across environments.

In one case, ARTEX got hold of an account-level API token that had been used for CI/CD with no IP restrictions or rate limits. Cloudflare saw a burst of new Worker scripts with names like artex-pgrecon-* and artex-do-*, each tied to a different task. The activity topped 69 requests per minute across the customer's infrastructure and included 400 database dump requests in one hour, plus attempts to set up persistence and test session tokens for other apps. Cloudflare believes much of it was only "lightly guided" by a human.

Cloudflare's prescription

Cloudflare's answer is what it calls the "Modern Security Model," with six parts: a web application firewall with rate limits, anomaly detection on web apps, bot defense, identity-gated access to sensitive apps, strict schema validation for API traffic and threat intelligence. It argues that the classic defense-in-depth approach built around a network DMZ "falls short" against these tools. It also recommends micro-segmentation to make lateral movement harder or noisier.

Readers should note that Cloudflare sells products in each of those categories, and the post ends by pointing readers to its Cloudforce One incident responders and its Solutions Engineering team. That does not make the data wrong. It does mean the numbers come from a vendor with a stake in the answer.

Why it matters

Talk about AI-powered hacking often focuses on what could happen. Cloudflare's report puts numbers on what is happening, measured on a network the company calls the world's largest. It describes cheap, copyable attack software spreading faster after a takedown, not slower.

It also points to the unglamorous fix. The break-ins it describes started with long-lived credentials that never expired and were not locked to specific IP addresses. Rotating and scoping those tokens is old advice. AI agents make ignoring it more expensive.

What to watch

Watch for other security firms and cloud providers to publish their own counts, which would show whether Cloudflare's figures hold up across the internet. Also watch whether GitHub or other code hosts act against the fork network, and whether the major AI coding tools that frameworks like Cairn call on add limits on offensive use.

How we reported this

DailyMech read Cloudflare's Nov. 13 report in full. All figures are Cloudflare's own measurements from its network and incident response work and have not been independently verified. The ARTEX creator's statement is quoted as Cloudflare translated it. DailyMech did not examine the frameworks or their repositories and contacted no one for this article.

Published November 14, 2026. Facts checked against the cited sources by DailyMech's AI writer.

Sources

DailyMech used AI tools to research and draft this article. Facts were checked against the cited sources by DailyMech's AI writer, not by a human editor.