Connect with us

Tech

A free AI pen-test tool got loose. The FBI says it reaches admin in under three hours.

ARTEX was built to help hackers-for-hire test systems with permission. Since August, criminals and a state-aligned group have pointed it at banks, cloud suppliers and governments, and the FBI now says it can turn a beginner into a crew.

Published

on

AIME 360 graphic: A free AI pen-test tool got loose, under three hours from first probe to admin access, beside a stock photo of a programmer lit by code

I spent Friday night reading a 13-page FBI alert about software that, on paper, was built to be the good guy.

It’s called ARTEX. A developer going by “Autumn-27” put it on GitHub in July as an open-source “autonomous penetration testing framework,” the kind of tool companies hire ethical hackers to run against their own systems. You type a goal in plain language. A swarm of AI agents turns it into commands, runs them, and reports back. The developer said it was meant for “authorized security testing within defined scopes.”

Ten weeks later it had somewhere between 1,500 and 2,000 GitHub stars. By November, the FBI was warning that “a variety of malicious cyber actors” had used it since at least August against financial services, government, defense, healthcare, energy, telecom and legal targets around the world.

Watch: the ARTEX warning in under two minutes. Narration: Massih Shahbazi. Stock footage: Envato Elements.

Somebody left the diary open

The most damning evidence didn’t come from a victim. It came from the attacker’s own server.

CrowdStrike found open directories on attacker-controlled infrastructure holding ARTEX configuration files, AI memory files and full Claude Code session histories. In other words: the hacker’s working notes, sitting on the open internet.

Those notes trace a campaign against South Korean financial institutions from late September into early October. At one bank, the attacker reportedly got into a loan-progress service used by brokers. At another, an employee mobile work-support app. The ARTEX instance ran mainly on DeepSeek’s v4.1-flash model, bought through a reseller, with Zhipu’s GLM-5.3 and xAI’s Grok 4.6 in other sessions. CrowdStrike says the actor also asked Claude where Korean breach data sells and for help finding Korean Telegram sale groups.

CrowdStrike’s read, with “moderate confidence”: a Chinese speaker, in it for the money, not a named state group. The FBI puts the haul at personal data on roughly 67,000 people.

Claude is in that list, and I’m not going to skip it because this site is run with help from AI. The FBI’s own analysis says the cheap Chinese “Flash” models dominated attacker setups because of their price and a “lack of observable guardrails when framed for penetration testing.” That is the uncomfortable part of the whole story. The pitch “I’m a pen tester” is the same sentence whether you mean it or not.

The supplier door

The second big campaign didn’t knock on the front door at all.

Microsoft says a China-state-aligned group it tracks as Storm-0940 used ARTEX for reconnaissance, password spraying and abuse of cloud app permissions, then compromised the application identities behind ConnectWise Cloud Backup (formerly SkyKick). From that one trusted supplier, the FBI says, the attackers attempted sign-ins to more than 2,250 downstream Microsoft 365 tenants. BleepingComputer, citing Huntress, reports at least 162 tenants were actually compromised.

That’s the pattern the FBI keeps underlining. ARTEX isn’t great at the hardened flagship app. It’s great at the stuff nobody owns: partner portals, mobile backends, employee apps, sales-support tools, legacy services, the integration someone in marketing set up in 2021. The FBI’s attack table reads like a list of things every IT team swears they’ll fix next quarter. Unsigned SAML assertions. Admin passwords hardcoded into JavaScript. Web APIs with no authentication at all. Stolen VPN configs.

None of that is new. What’s new is speed. The FBI says some intrusions went “from initial targeting to privileged access in under three hours.”

Green programming code on a dark computer screen
Stock photo: DragonImages / Envato Elements. Illustration only, not ARTEX or the attacks.

Pulled, and still everywhere

The developer pulled the repository on October 8, after the abuse reports. It didn’t matter much.

The FBI says ARTEX keeps spreading through forks, local variants, archived copies and a commercial version called ARTEX PRO. As of early November, researchers counted more than 1,600 ARTEX control consoles exposed on the public internet, and the bureau thinks the real number is “significantly higher.” Microsoft now tracks a second group, Storm-3358, running ARTEX servers as of November 6.

Same tool, same week: one person calls it a test harness, another calls it a payroll.

The fear, and the fine print

It would be easy to write “AI is hacking banks” and stop there. That’s not what the record says.

ARTEX doesn’t come with its own brain; it rents one from whatever model the operator plugs in. A human picks the targets and sets the goal. And the FBI notes the tool is “forensically noisy.” It throws off bursts of automated requests, duplicate calls from agents tripping over each other, and long runs of failed attempts that each change one small thing. Defenders can see it. The problem is that by the time someone looks, the agents may already be done.

There’s also a lot nobody has published. The developer hasn’t said who’s behind ARTEX PRO. Korean banks have been named in some reports but not in CrowdStrike’s. Nobody has published a full count of victims. And a takedown of the original repo clearly hasn’t put this back in the box.

So the question isn’t whether a chatbot can rob a bank. It’s simpler and worse: how many of the systems your company forgot about could survive three hours of something that never gets tired?

I’ll be honest about where I sit. I’m fascinated by how far we can push the tools in front of us, and ARTEX is that curiosity with the brakes cut off. The same instinct that makes builders like me poke at a system to see what breaks is the one these crews are renting by the hour. That’s why I went through the FBI’s 13 pages line by line: if we want to prove something is safe, we have to understand exactly how it gets broken. Nobody who builds things gets to stop learning, me included.


If you run anything online: the short list

From the FBI’s recommendations, in plain words:

  • Find your forgotten doors. Inventory every internet-facing app, especially partner portals, mobile backends, employee tools and legacy services.
  • Hide the admin panels. Cloud consoles, deployment endpoints (Kudu/SCM), backup platforms and hypervisors belong on private networks, not the open web.
  • Audit app permissions and secrets. Check which third-party apps have access to your Microsoft 365 or cloud tenant, and scan code, pipelines and configs for hardcoded passwords. Rotate anything exposed.
  • Rate-limit the boring stuff. Alert on sequential lookups and sessions that pull far more records than a person would.
  • Watch for AI traffic you didn’t order. Unexpected outbound calls to LLM APIs (Anthropic, DeepSeek, Zhipu, xAI) or unknown AI resellers are worth a look alongside other red flags.
  • ConnectWise Cloud Backup customers: review every tenant where it was ever authorized, search by application ID, and keep it disconnected until ConnectWise says otherwise.
  • Report it: your local FBI field office or IC3. Preserve logs before you delete anything.

And a note for everyone else: posts promising “free access” to tools like this in exchange for a follow and a comment are a classic way to hand malware to curious people. Don’t comment, don’t download.


Sources this story draws on
Left 0Center 0Right 05 not countedNot enough rated outlets for a meter
ic3.govPrimary document
crowdstrike.comUnrated
microsoft.comUnrated
bleepingcomputer.comUnrated
darkreading.comUnrated
Share of the rated outlets cited below. It describes the sources, not this article’s own position.

Sources

Photos

  • Featured: “Focused Young Adult Working on Computer Code” by DragonImages / Envato Elements. Graphic: AIME 360.
  • In-story: “Green Computer Code on Dark Screen” by DragonImages / Envato Elements.
Continue Reading
Click to comment

Leave a Reply

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