
ARTEX tracker: campaigns, forks and exposed consoles of the AI hacking agent
By Himanshu Sakre·
A living, dated log of the open-source ARTEX framework that attackers use to automate intrusions: what the FBI, researchers and defenders have documented since July 2026, with what is confirmed and what is single-sourced.
ARTEX is the first widely documented case of an open-source AI hacking agent moving from a security-testing project to a standard attacker tool within weeks. This page is the running record. It is updated through our corrections loop whenever a new fact appears, so check the dates. Every entry names its source, and single-source claims are marked.
What it is
The FBI describes ARTEX as "an open-source, large language model (LLM)-driven agentic platform" that threat actors have weaponized. It is written in Go, uses specialized agents to turn goals stated in natural language into commands, and ships with no model of its own. It connects to outside providers, including Anthropic Claude and OpenAI GPT, and to DeepSeek, GLM, Grok and Kimi. The Cloud Security Alliance (CSA) adds that it is licensed under AGPL-3.0 and has a Next.js console and a PostgreSQL backend.
Timeline
- July 2026: The developer, using the handle Autumn-27, releases ARTEX on GitHub. The FBI estimates 1,500 to 2,000 stars within 10 weeks.
- Late August 2026: CSA dates the earliest clusters of malicious use here. The FBI says abuse has run since at least August.
- October 3 to November 5, 2026: Amazon's MadPot sensors collect ARTEX traffic, per CSA.
- October 8, 2026: After reports of abuse, the developer says the project will go closed-source and removes the GitHub repository and official Docker image.
- November 5, 2026: CERT-EU warns it has seen ARTEX-assisted attacks against organizations it oversees, per The Register.
- Early November 2026: Researchers find more than 1,600 exposed ARTEX consoles.
- November 13, 2026: The FBI issues FLASH-20261113-001.
- November 16, 2026: CSA publishes a research note on ARTEX and its forks.
- November 19, 2026: The Register, citing Politico, reports ARTEX was reportedly used against European Commission systems. We cover that in our report on the European Commission attack.
Documented campaigns
- South Korean financial institutions: Financially motivated attacks led to exfiltration of personal information of about 67,000 individuals (FBI).
- ConnectWise Cloud Backup supply chain: Attempted sign-ins to more than 2,250 downstream tenants (FBI).
- VAPOR PANDA: CrowdStrike reported activity against more than 200 organizations in 14 countries, mainly using ARTEX with DeepSeek (CSA summary).
- UNC7228: A cluster that overlaps with Microsoft's Storm-0940 (CSA summary of Google's reporting).
- European Commission: Reported only, via Politico's sources, as of November 19, 2026.
Speed, as measured
The FBI says some intrusions progressed from initial targeting to privileged access in under three hours. CSA cites Google's threat intelligence group at about four hours on average from first access to deep access, and 15 minutes in one case. Wiz observed sessions lasting 24 to 30 hours, with new source addresses appearing within minutes of each block. These measure different things and come from different vendors.
Forks and exposed infrastructure
Withdrawal removed the official download and little else. CSA lists ScopeWeaver (an English-localized rebrand), ARTEX PRO (a commercial edition), and consoles branded "ARTEX MX" and "ARTEX v2 Agent Security Workbench." It also lists AntiFox, a fork of ARTEX 0.1.1 that reached version 2.0 and was used in AWS credential-theft intrusions per Ostorlab. Active forks have reached v0.5.1 while the last upstream release was v0.3.15.
“faster than a single maintainer or platform can withdraw it”
Counts of exposed consoles differ by method. Censys counted 1,465 hosts in early November, about 815 distinct-looking deployments. Ostorlab gives a range of roughly 800 to 1,650. CSA says these are not directly comparable and that some consoles may belong to authorized testers. About 10.6% of Censys hosts also ran other threat tooling such as Cobalt Strike.
Trojanized copies are a second risk: CSA says 637 GitHub repositories still referenced a trojanized Docker image on November 5, and about 34 Docker Hub namespaces published "artex" images with roughly 60,000 combined pulls. The reporting does not say how many pulls were malicious.
Which models power it
The FBI says analysis of attacker-run instances shows Chinese models are most common, with DeepSeek Flash family models dominating because of cost, capability and "lack of observable guardrails when framed for penetration testing." That is a claim about attacker choice, not about any vendor's safeguards, and the bureau does not publish the sample. For where frontier labs are restricting cyber-capable models, see our tracker of restricted-access cyber AI models.
What defenders are told
The FBI's guidance targets internet-exposed applications, self-service portals, cloud infrastructure and partner systems. CSA's immediate list is to remove default and hardcoded credentials, rotate exposed secrets, close unauthenticated admin interfaces and inference endpoints, patch internet-facing software, and hunt identity-provider logs for password-spray and account-recovery abuse. It also warns that manual IP blocking may lag agents that change addresses within minutes.
Open questions
- Whether the European Commission incident involved data theft is not established.
- No public figure separates ARTEX intrusions from all AI-assisted intrusions.
- Attribution beyond the FBI's cluster descriptions is mostly vendor reporting.
How to read a single number
The 12,000 fork events and the 49 distinct copies are not in conflict. CSA says the first is raw GitHub activity and the second is a curated count of distinguishable lineages. Likewise the 1,600 exposed consoles, the 1,465 Censys hosts and the 815 distinct-looking deployments measure the same thing at different levels of cleaning. When this page quotes a count, the method is named so later figures can be compared properly.
Our take
The lesson is the commodity effect. The tool did not need to be good, only free and fast. We will update this page when the Commission, CERT-EU or the FBI adds a fact, and we will mark any number that cannot be traced to a named source.
Frequently asked questions
What is ARTEX?
An open-source, LLM-driven agentic penetration testing framework written in Go, released on GitHub in July 2026. The FBI says threat actors have weaponized it to automate intrusions at machine speed, from reconnaissance through data exfiltration. It has no model of its own.
Is ARTEX still available after the developer removed it?
Yes. The developer removed the repository on October 8, 2026, but forks, archives, localized versions and a commercial edition called ARTEX PRO remain. The Cloud Security Alliance counts more than 12,000 fork events recorded by Wiz.
How many ARTEX consoles are exposed to the internet?
The FBI cites more than 1,600 exposed consoles found in early November 2026. Censys counted 1,465 hosts, about 815 distinct deployments, and Ostorlab gives roughly 800 to 1,650. Some may belong to authorized testers.
Which AI models does ARTEX use?
It connects to outside providers, including Claude and GPT, and to DeepSeek, GLM, Grok and Kimi. The FBI says Chinese models are most common in attacker-run instances, with DeepSeek Flash family models dominating, based on its analysis.
How fast are ARTEX attacks?
The FBI says some intrusions reached privileged access in under three hours. The Cloud Security Alliance, citing Google's threat intelligence group, gives about four hours on average and 15 minutes in one case. The measures come from different vendors.
Was the European Commission attacked with ARTEX?
Reportedly. The Register, citing Politico's sources, says defenders detected ARTEX in the Commission campaign. The Commission has confirmed a detected cyber campaign but not named the tool, and whether data was stolen is not established.
Sources
What each one is, and whose it is.
- 1
Threat Overview of ARTEX Autonomous Penetration Testing Framework (FLASH-20261113-001), FBI Internet Crime Complaint Center (November 13, 2026)
OtherIndependent of the vendor - 2
ARTEX and Its Forks: Agentic Pentest Frameworks as Attack Infrastructure, Cloud Security Alliance AI Safety Initiative (November 16, 2026)
OtherIndependent of the vendor - 3
European Commission scrambles to contain AI-assisted cyberattack, The Register (November 19, 2026)
Press reportIndependent of the vendor