FBI warns: AI hacking tool ARTEX breaks into companies in under three hours

The FBI is warning about ARTEX, a freely available AI tool for security testing. Since August, criminals and state-linked groups have been abusing it for attacks around the world. The program lets AI agents find and exploit vulnerabilities on their own, and some attackers reached admin rights in less than three hours.

Industry Nov 16, 2026 7 min read Source: Bing News (nl-NL)
Illustrative image, AI-generated

What's going on

On November 13, 2026, the FBI issued what it calls a FLASH alert, a short-notice technical advisory for companies and government agencies. This one is about ARTEX, a freely available program for automated penetration testing. In a penetration test, security professionals attack a company's systems with its permission to find weaknesses before real attackers do.

According to the FBI, a range of criminal and state-linked groups have been using this very tool since at least August 2026 for real attacks on organizations worldwide. The FBI lists financial services firms, government agencies, technology and defense companies, semiconductor makers, energy providers, healthcare, manufacturing, telecoms and law firms among the targets. Italian outlet HWUpgrade reports that the 13-page document carries the identifier FLASH-20261113-001.

The speed is what's alarming. The FBI says that in some intrusions, attackers went from first scouting a target to privileged access, meaning admin-level rights, in less than three hours.

How ARTEX works

ARTEX was released in July 2026 by a GitHub user called "Autumn-27". It's written in Go and ships with a web interface built on the Next.js framework. At its core it's a command-and-control (C2) platform, the hub from which attacks are directed. What's new is that this hub doesn't just pass along commands from humans. It coordinates several AI agents, programs that plan and carry out steps on their own.

HWUpgrade says specialized agents split up the work: reconnaissance, vulnerability discovery, exploitation, privilege escalation and finally data exfiltration. All of it happens with little human involvement. Taiwanese tech site iThome puts it this way: the user describes the goal in plain language, and the agents turn that into the concrete commands a human would otherwise have to type one by one.

ARTEX doesn't have its own AI model. It plugs into external large language models (LLMs), from providers like Anthropic (Claude) and OpenAI (GPT) as well as open or Chinese models such as DeepSeek, GLM, Grok and Kimi. The FBI writes:

"Analysis of attacker-controlled ARTEX instances shows that Chinese models are the most widely used, with models from the DeepSeek Flash family dominating, due to their combination of capability, relatively low cost and the absence of noticeable safeguards when the request is framed as a penetration test."

Those safeguards, or guardrails, are the built-in restrictions AI providers use to stop their models from helping with harmful tasks. With some models, they apparently stop working once an attack is presented as an authorized security test.

The known attacks

The FBI and other reports point to several specific incidents:

  • ConnectWise Cloud Backup: In October, attackers hit the ConnectWise backup service (formerly SkyKick Cloud Backup) in a supply chain attack. Instead of breaking into a target directly, a supply chain attack goes after a service provider that many targets trust. According to HWUpgrade, the China-linked group Storm-0940 hijacked the service's application identities and used ARTEX to try to log into more than 2,250 end-customer Microsoft 365 environments. Security firm Huntress found that it worked in at least 162 cases. iThome stresses that the 2,250 figure doesn't mean all of them were successfully breached.
  • South Korean banks: HWUpgrade reports that a Chinese-speaking actor combined ARTEX with Claude Code agents to attack institutions including Shinhan Bank, KB Kookmin Bank and Hana Bank, stealing personal data on around 67,000 people. In a separate case, South Korean police arrested seven suspects linked to intrusions at more than 90 organizations, including clinics, schools and businesses, and the theft of about 8.2 million records.
  • Vapor Panda: According to HWUpgrade, CrowdStrike documented a campaign by the espionage group Vapor Panda against more than 200 organizations in 14 countries, with at least nine confirmed compromises.
  • 91APP in Taiwan: iThome reports that e-commerce provider 91APP disclosed on October 22 that an order database on its Commerce Cloud platform had been breached. The attackers exploited an unpatched flaw in Appsmith, a tool the company used internally, then pulled a database key from its GitLab code platform. Investigators found traces of ARTEX.

Why ARTEX is so dangerous

The FBI sees several risks. First, ARTEX automates the tedious phases of an attack: probing systems and exploiting vulnerabilities. That lets even a small group run several campaigns at once. Second, it lowers the barrier to entry. The FBI says even people with little expertise can run complex operations against multiple targets at the same time.

Third, ARTEX is especially good at finding neglected systems. These aren't a company's flagship applications. They're partner and supplier portals, backend servers for mobile apps, internal tools for staff and sales teams, legacy systems and third-party integrations. They're often connected to internal databases but get less attention.

Fourth, ARTEX can map trust relationships between companies and use weaker suppliers or CI/CD pipelines, the automated workflows that build and ship software, as stepping stones. The FBI says attackers even tampered with iOS app distribution infrastructure this way, using exposed cloud credentials. iThome adds that attackers use ARTEX to validate stolen Azure keys, search for reachable cloud resources and change OAuth permissions, the grants that let apps access accounts.

ARTEX does generate a lot of noisy traffic that can trigger alerts. But the FBI says it moves so fast that defenders have little time to respond before attackers reach their goal.

Deleted, but not gone

Autumn-27 took down the original GitHub repository on October 8, 2026, citing widespread abuse. The developer insists they have nothing to do with the attacks and says the tool was only meant for authorized testing. That hasn't stopped it from spreading. Forks (independent copies of the code), archived copies, localized variants like ScopeWeaver and a commercial version called ARTEX PRO are all in circulation.

Censys, a search engine for internet-connected devices, found that some ARTEX servers even had open directories holding additional attack tools. According to figures cited by the FBI, more than 1,600 ARTEX admin interfaces were publicly reachable as of early November, and that doesn't count installations behind firewalls. Infosecurity Magazine reports that Censys concludes the real level of activity far exceeds the known campaigns.

What the FBI recommends

The FBI provides indicators of compromise, such as IP addresses and command servers, but warns that these change often. Above all, it wants organizations to harden their internet-facing applications, especially partner portals, legacy systems and third-party integrations. Access to cloud management consoles should be restricted and app permissions reviewed. iThome says that if you use Microsoft 365 and Microsoft Entra, you should check sign-in logs for service accounts and OAuth grants. For critical APIs, the FBI recommends rate limiting and monitoring for sudden request spikes, unexpected credential changes and large data transfers out of the network. Security researcher MalwareMustDie has also published an open package of detection signatures along with a playbook.

The FBI is also asking for tips: infrastructure traces, sign-in and access logs, API activity, signs of the AI infrastructure used, cloud and container activity, and evidence of exploitation and persistence.

The bigger picture

ARTEX is one of the first well-documented cases of agentic AI doing more than helping programmers. Here it speeds up attacks at scale. The tool doesn't necessarily bring new attack techniques. It just chains known steps together much faster and with less expertise. That shifts the balance: companies now have to protect every forgotten interface, not just their showcase systems.

The role of AI models stands out. Because an open toolkit can plug into any model, the safeguards of individual providers lose much of their effect. Attackers simply switch to cheap models with weak guardrails, as the FBI has seen with DeepSeek Flash. At the same time, the South Korea case shows that Western tools like Claude Code were also involved. The debate over open-source attack tools is an old one. ARTEX shows that deleting the original does little once the code is out there. It's still unclear how many breaches actually trace back to ARTEX, and even the count of reachable consoles is considered a lower bound.

Related videos

Risky Business (859): ARTEX damage approaches crisis levels Risky Business Media · YouTube

Sources

#cybersecurity#ai agents#fbi#hacking#deepseek
Read the original source

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.