Skip to content

Threat Actors

FBI says ARTEX AI hacking tooling persists after GitHub removal

The FBI says some ARTEX intrusions reached privileged access in under three hours.

https://breach-signal.com/threat-actors/fbi-artex-ai-tool-persists-github-removal/

Share

This is a new development in a story we reported earlier: Microsoft attributes ConnectWise supply chain compromise to China-state aligned Storm-0940 using ARTEX AI.

Key facts

  • Developer Autumn-27 published ARTEX on GitHub in July 2026 and took the repository down on Oct. 8 after abuse reports.
  • Localized versions, community forks, archived copies and a paid edition called ARTEX PRO kept the framework available after removal.
  • ARTEX agents each require calls to an external language model API because the Go-based framework has no model of its own.
  • Microsoft says Storm-0940 used ARTEX against more than 2,250 ConnectWise Microsoft 365 tenants, and most sign-in attempts failed.

The FBI warned on Nov. 13 that hackers are using ARTEX, an AI-powered penetration testing framework, to automate cyber attacks worldwide, and that the tooling remains reachable through GitHub forks, archived copies and a paid edition despite the original repository’s removal, with some intrusions reaching privileged access in under three hours.

ARTEX is a Go-based multi-agent system, meaning it coordinates several AI agents, with no built-in model of its own. Users state plain-language goals, and AI agents plan and run attack steps on a shared graph. Each agent requires calls to an external language model API, and the framework can integrate Anthropic Claude, OpenAI GPT, DeepSeek, GLM, Grok and Kimi.

Developer Autumn-27 published ARTEX on GitHub in July 2026 and took the repository down on Oct. 8 after abuse reports. The FBI says malicious use dates back to at least August 2026, before the removal. Localized versions, community forks, archived copies and a paid edition called ARTEX PRO kept the framework available after the takedown.

The FBI says ARTEX has been used against financial services, government, technology, defense, communications and law firms. CrowdStrike said Vapor Panda, a Chinese cyberespionage group, ran automated ARTEX attacks against more than 200 organizations in 14 countries and compromised at least nine, with AI agents performing reconnaissance, exploiting vulnerabilities, deploying web shells and stealing credentials.

CrowdStrike reported on Oct. 7 that an unattributed likely Chinese-speaking financially motivated actor used ARTEX against South Korean financial firms and found exposed infrastructure with ARTEX configuration files and Claude Code session histories. The FBI puts the theft from that campaign at about 67,000 people; TechCrunch reports at least 68,000. The actor used ARTEX with Claude Code agents against Shinhan Bank, KB Kookmin Bank and Hana Bank. South Korean police arrested seven suspects tied to attacks on more than 90 organizations, with ARTEX reportedly used in most intrusions and about 8.2 million records stolen.

Storm-3358, a financially motivated cluster based in Cambodia, started using ARTEX against financial firms in October, and Microsoft tracks it; the FBI says its infrastructure was active on Nov. 6. Vendors have not published proof linking ARTEX users to one another beyond their shared use of the framework.

The FBI says Storm-0940, a China-state-aligned actor active since 2021, used ARTEX against more than 2,250 ConnectWise Microsoft 365 tenants, and most sign-in attempts failed. Its password-spray campaign peaked at about 2,900 failed sign-ins per minute, and the company says Storm-0940 achieved persistent privileged cloud access at many organizations, though a successful authentication does not by itself prove data theft. The FBI says a ConnectWise supply-chain compromise reached more than 2,250 downstream tenants, but Microsoft says most attempts failed.

ConnectWise suspended several services, and its Nov. 4 advisory says Cloud Backup is partially restored in phases while partners must re-authorize each Microsoft 365 tenant to a replacement app. The ConnectWise attack used an application credential rather than a person’s password, so multifactor authentication did not challenge the stolen credential. Huntress said it identified at least 162 compromised ConnectWise tenants.

Censys counted 1,465 hosts running exposed ARTEX consoles on Nov. 4, with 726 in the United States, 387 in Hong Kong and 177 in China. Check Point Research found 2,557 systems hosting ARTEX panels, of which 1,608 were live, while the FBI estimates public exposure at more than 1,600 consoles in early November.

Roughly 650 exposed consoles occupy two identically configured hosting blocks, leaving about 815 distinct deployments, while 155 hosts also ran Cobalt Strike. AntiFox, a fork of ARTEX version 0.1.1, remains on GitHub as version 2.0, and Arctic Wolf linked AntiFox interfaces on Alibaba Cloud to AWS secret harvesting. Badbyte reported a copied ARTEX repository linked to a booby-trapped Docker image granting root shell access with a secret trigger, and 62 internet-facing consoles served the backdoored build between Nov. 2 and 4.

The FBI says attacker-run ARTEX instances mostly use Chinese models, led by DeepSeek’s Flash family for capability balance, low cost and weak guardrails. Check Point found American models on only three of ten analyzed ARTEX instances, with guardrails triggered only on Anthropic Opus 5.5 before the actor moved to more permissive Chinese models. CrowdStrike located Claude Code session histories and DeepSeek as the main ARTEX backend in directories exposed by the South Korea actor. Operators can route ARTEX model requests through self-hosted gateway and reseller software.

Amazon CISO CJ Moses said ARTEX techniques in observed campaigns were not new, based on decoy sensors capturing agent memory between Oct. 3 and Nov. 5. The company said with moderate confidence that separate Chinese-nexus groups were behind campaigns where ARTEX agents hunted, validated and pooled exposed model-provider keys. It also observed a second campaign attempting identity checks at two cryptocurrency exchanges using real-time deepfake video and forged IDs, and the vendor’s server rejected every deepfake attempt.

Amazon advises watching for unexpected AI spend because model-access logs and AI spending can show an agent at work after code is copied beyond a developer’s reach. A 2025 Commerce Department evaluation found older DeepSeek models fell behind leading U.S. models on nearly every benchmark and were more vulnerable to agent hijacking. TechRepublic’s report said DeepSeek agreed with public jailbreak prompts at rates of 95 percent to 100 percent, while U.S. models showed rates of 5 percent to 12 percent.

The Data Security Law of 2021, effective Sept. 1, 2021, requires Chinese government approval before organizations send mainland-stored data to foreign courts or law enforcement. China’s Cybersecurity Law of 2017 took effect June 1, 2017, was amended in 2025 and requires network operators to provide technical assistance to public security organs under Article 28. Public reporting has not surfaced any independent audit of what Chinese model providers have handed to authorities.

The FBI recommends inventorying every internet-facing application and keeping cloud administration, Kudu, SCM, hypervisor and backup interfaces off the public internet. It instructs organizations to use least privilege for multi-tenant application permissions and consent grants, to monitor credential additions, role changes and application-only behavior, and warns against deleting enterprise apps first because that destroys evidence. Organizations with affected ConnectWise apps should review tenants by application ID, look back to at least Oct. 1, and keep connections disabled until vendor restoration guidance.

Victims should report ARTEX activity to a local field office or the Internet Crime Complaint Center at ic3.gov, the FBI says. It cautions that cloud IP addresses get reassigned, so indicators should be checked against current telemetry before any blocking. Its 13-page advisory lists indicators such as suspicious network activity, command-and-control infrastructure and IP addresses, and it flags API calls to Anthropic, DeepSeek, GLM and xAI as indicators worth investigating, not automatic blocks. MalwareMustDie published an open-source ARTEX detection signature pack and a playbook that sorts indicators by confidence level.

Barracuda disclosed an ARTEX-linked breach in which financially motivated criminals used an exposed third-party credential to access customer credential stores. It could not rule out access because its logs lacked detail, so it changed all customer credentials. “Defenders who rely on spotting malicious activity and then responding to it will often arrive after the attacker has finished,” wrote Arve Kjoelen, the company’s CISO.

Android Headlines reports that ARTEX usually goes after forgotten partner portals and weak passwords or accounts missing multifactor authentication, not rare software flaws, and that rate-limiting APIs, canceling old vendor logins and shutting public admin pages can stop automated bots before they spread. An Aviatrix executive, Matt Snyder, told Dark Reading that many ARTEX deployments have weak operational security despite the framework’s sophistication.

AndroGuider says no publicly available FBI advisory, press release or court filing independently confirms an Artex AI tool matching that description. The outlet warns against presenting the story as fact without corroboration.

Was this helpful?

Join the discussion

Please read our Comment Policy before commenting. First-time comments are held for moderation. By posting a comment you agree to our Terms of Use and Privacy Policy, which explain how your name, email address and comment are used.

Your email address will not be published. Required fields are marked *