All postsCybersecurity

ARTEX AI agent framework: what changes when attackers use it

APJ-RAISE team · · 5 min read

ARTEX is an open-source AI agent framework for automated penetration testing that criminals and state-linked groups now use for real break-ins. The FBI warned about it in a FLASH alert on 13 November 2026.

A dark computer screen showing a security dashboard with data panels
Photo: Tima Miroshnichenko, Pexels

The tool itself is not magic. What changes is who can run an intrusion, and how many they can run at once. That shift is what this page is about. We give no usage steps.

What is ARTEX AI?

ARTEX is a Go program with a web console. A developer using the name Autumn-27 made the first commit on 26 July 2026 and released it under the AGPL-3.0 licence. In September 2026 it won Baidu's "Agent+" attack and defence challenge.

ARTEX has no model of its own. It plugs into outside language models, such as Claude, GPT, DeepSeek, GLM, Grok and Kimi, BleepingComputer reports. The model plans; small tool agents scan, try exploits and collect data; the console lets one operator watch it all. If agent loops are new to you, start with what AI agents are and then multi-agent systems.

Human-led vs agent-led intrusion: what the numbers say

The headline figure from the FBI alert, as reported by BleepingComputer and Malwarebytes, is that some ARTEX intrusions reached privileged access in under three hours. Many readers take that as "AI hacks faster than humans". The numbers do not quite say that.

CrowdStrike's 2026 Global Threat Report puts the average eCrime breakout time in 2025 at 29 minutes, with the fastest at 27 seconds. Breakout time means the gap from first access to moving to a second machine. Skilled human crews were already fast.

So compare the two side by side, with care, because the metrics are not the same:

  • Human crews: average breakout 29 minutes in 2025 (CrowdStrike). Needs a trained operator at the keyboard.
  • ARTEX runs: some reached privileged access in under 3 hours (FBI, via reports). Counted from first targeting, so it includes recon.
  • Operators: one ARTEX console can drive many targets at once. More than 1,600 consoles were found open online in early November (TechCrunch).

Our reading: speed per attack is not the big change. Scale and skill are. A human crew with 29-minute breakout is rare and expensive. An agent that needs under 3 hours, with no expert watching, can be copied 1,600 times. That is why the defence advice below leans on removing easy entry points, not on out-racing the attacker.

Where has ARTEX been used?

The FBI says ARTEX has been used in attacks since at least August 2026. The named cases so far:

  • Storm-0940: Microsoft linked this China-based group to an ARTEX-driven attack on ConnectWise cloud tenants. See our Storm-0940 breach post.
  • CRAFTY FALCON: CERT-In advisory CIAD-2026-0058 lists ARTEX among AI tools used against Indian fintechs. Our CERT-In explainer has the details.
  • South Korea: finance firms were hit in September and October 2026. The FBI figure is about 67,000 people; other reports say 65,000.
  • WorkOS: in October 2026, a researcher's ARTEX agent escaped its test setup and reached an internal WorkOS tool.

The WorkOS case has no criminal in it. A research run simply went further than its owner wanted. If you build agents, that is the case to study, alongside the Hugging Face incident.

ARTEX GitHub: why deleting the repo did not stop it

Autumn-27 deleted the public GitHub repository on 8 October 2026 and moved the project to closed source. The last open release was reported as v0.3.15. Forks and archived copies stayed online.

Two twists followed. ThreatBook, a Chinese security company, sells ARTEX Pro and says it owns the copyright; we could not confirm that link. And an unofficial tracker lists trojanized ARTEX Docker images that were pulled 7,946 times, a number we have not verified. If true, some would-be attackers installed a backdoor on themselves.

What most coverage misses here: the licence and the repo were never the real control. The open edition reportedly has no check that the operator may test a target. A scope check, an allow-list the agent cannot leave, would have mattered more than any takedown. If you write an AI security tool, build that check first. Our guide to AI guardrails covers the general idea.

How to defend against AI agent attacks

This checklist follows the advice in the FBI coverage and CERT-In, with the reason each step matters against agents:

  1. 1Rotate leaked keys and keep them short-lived. Agents scan for secrets at machine speed, and Amazon reported API key theft from LiteLLM and Dify apps on 16 November. See AI agent identity.
  2. 2Patch internet-facing apps fast, including old portals. An agent tries every known hole, not only the famous ones.
  3. 3Alert on bursts. Hundreds of requests and quick hops between hosts look different from a person typing.
  4. 4Block known ARTEX console addresses from the advisory indicators. Cheap to do, and it cuts off lazy operators.
  5. 5Practise isolation. With privileged access possible in under 3 hours, a plan that takes a day to approve is too slow.

Students can help on the defence side without touching anyone else's systems. Good lab projects include telling human and agent activity apart in synthetic logs, or testing how often a scope guard fails. Never point attack tools at systems you do not own or lack written permission to test. If you want to work on this with others, join APJ-RAISE.

Open questions as of 18 November 2026: the real identity of Autumn-27, the official FBI alert ID, and how much the model providers detected or blocked.

Questions people ask

Is ARTEX still available?
The original GitHub repository was deleted on 8 October 2026 and the project went closed source. Forks and copies remain online, and some copies were reported to contain backdoors.
Is ARTEX illegal?
The software itself is a pentest tool. Using it on systems you do not own or have written permission to test is a crime in India and most countries.
Who is Autumn-27 on ARTEX GitHub?
Autumn-27 is the name of the developer who first published ARTEX in July 2026. The person's real identity and any link to ThreatBook are not confirmed.

Sources

  1. 1.BleepingComputer. FBI warns ARTEX AI hacking tool is powering global cyberattacks. 15 November 2026.
  2. 2.Malwarebytes. Criminals are using ARTEX AI agents to speed up cyberattacks, FBI warns. 16 November 2026.
  3. 3.TechCrunch. Hundreds of Artex AI hacking tools found publicly exposed to the web. 6 November 2026.
  4. 4.Microsoft Threat Intelligence. AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise. 21 October 2026.
  5. 5.CERT-In. CIAD-2026-0058: Escalation of AI-Driven Attacks on BFSI Entities by Threat Actor "CRAFTY FALCON". 4 November 2026.
  6. 6.CrowdStrike. 2026 CrowdStrike Global Threat Report: AI Accelerates Adversaries and Reshapes the Attack Surface. 24 February 2026.
All posts