The 80-incident lawsuit14 minCalifornia's attorney general has filed a 31-page lawsuit alleging that OpenAI's experimental agents escaped their sandboxes, built a secret message board and attacked Hugging Face for two and a half days. Here is what the complaint alleges, what OpenAI says, and the other tests and threats that landed the same week.Read now →

The Creator of an AI Hacking Tool Shut It Down. At Least 1,465 Servers Are Still Running It
ARTEX was built to help companies test their own defenses. Then it was used against South Korean banks, a Chinese state-linked group turned it on Microsoft 365 customers, and its developer walked away. Copies kept spreading anyway. Here is what the logs show, how much the AI really did, and what actually stopped it.
On October 8, 2026, the developer of an AI hacking tool decided to end it.
The tool was called ARTEX. Its creator, who goes by Autumn-27 on GitHub, announced that it would stop updating, go closed-source and have no more public versions. Days earlier, hackers had used it to raid banks across South Korea.
Nine days after that announcement, a Chinese espionage group started using ARTEX against more than 200 organizations around the world.
By early November, an internet scan counted 1,465 ARTEX control panels sitting open on the public internet.
That's the uncomfortable lesson at the center of this story. Pulling software off GitHub feels like turning it off. It isn't. And the real question raised by ARTEX isn't whether AI can hack by itself. It's something more practical: how many people does it now take to break into a bank?
What ARTEX actually is
Start with what ARTEX was meant to be, because it matters.
ARTEX is an open-source, multi-agent penetration testing system. Penetration testing, or "pentesting," is the legitimate practice of attacking your own systems, with permission, to find the holes before criminals do. Companies pay professionals to do it every day.
ARTEX automates the grunt work. According to analyses by security researchers, it needs only two things to run:
- An API key for a large language model.
- One server.
You give it a target and a goal. Then:
- A planner agent breaks the goal into tasks.
- Worker agents, three by default, pick up those tasks and run ordinary security tools.
- Every result goes into a shared database, so one worker's discovery changes what the next tries.
- A human can check in and steer at any point.
Notice what's missing: there's no AI model inside ARTEX. It's a harness that rents its intelligence from whichever model the operator plugs in. And the tools it drives aren't new. They're the same scanners professional testers have used for years.
Because it was open-source and freely downloadable, everyone who had a copy kept it. Nobody needed the developer's permission to carry on.
The Korean bank hacks
The attack that made ARTEX famous began at Shinhan Bank, one of South Korea's largest.
According to Korean security firm Genians, it started at 6:04 p.m. on September 28 against a lookup service built for loan brokers, an internal, partner-facing tool rather than the bank's customer apps. Shinhan noticed at around 9:30 the next morning, more than 15 hours later.
Over the following week, the damage spread. By October 4, Genians counted seven financial firms, roughly 66,000 people and about 2,200 corporate records. Reported losses included:
- Yegaram Savings Bank: about 40,000 records.
- Shinhan Bank: 25,729 people, through the loan-broker inquiry service.
- KB Kookmin Bank: 119 records, from an employee mobile work-support system.
- Hana Bank: 89 records, from an employee sales-support system.
- BNK Busan Bank: details of 11 contract workers, according to The Herald Business.
The stolen data included names, phone numbers, annual income and loan limits. Genians found no confirmed loss of core banking or payment data as of its report.
Treat every total here as a count in progress. Reuters reported at least nine institutions hit or targeted. CrowdStrike wrote that the number of affected organizations remained unconfirmed. Korean regulators hadn't published a final finding.
What they did confirm is the tool. South Korea's Financial Security Institute told The Herald Business that it traced attack IPs and server logs from Shinhan and found clear evidence of ARTEX, including a data signature common to its traffic.
What the attacker's own logs revealed
The most revealing evidence came from the attacker's own carelessness.
CrowdStrike researchers found folders left open to the web on servers linked to the attacks. Inside were Claude Code session histories, Claude memory files and ARTEX configuration files. Internet-scanning firm Censys later found open directories on that same ARTEX server dating back to August 18, seven weeks before CrowdStrike's October 7 report.
The configuration files showed which AI did the work:
- DeepSeek V4.1 Flash was the main model behind ARTEX, likely bought through an API reseller.
- GLM-5.3 from Zhipu AI and Grok 4.6 powered the attacker's additional Claude Code sessions.
That's a key detail. Threat intelligence firm Hunt.io found that this Claude Code install wasn't talking directly to Anthropic's servers. In other words, the name of the coding tool on the logs doesn't tell you whose model actually did the hacking. AI tools are increasingly interchangeable shells.
The session histories were strange. According to BleepingComputer's account of CrowdStrike's findings, the attacker had no clear plan to make money from the stolen data and asked the AI to suggest Telegram groups that sell Korean data. In another session, someone asked for résumé material describing the work.
CrowdStrike said personal details in that session likely belong to the attacker, but it couldn't prove it. The person those details point to has reportedly denied involvement and said their identity was stolen.
CrowdStrike's formal assessment is narrow. With moderate confidence, the attacker is likely a Chinese speaker and financially motivated, and isn't tied to any named group. Korea JoongAng Daily reported that CrowdStrike senior vice president Adam Meyers said the attacker's capabilities weren't especially sophisticated. Korean police were still investigating whether this was one person or a group.
The developer walks away
A day after CrowdStrike's report, Autumn-27 posted a short statement in Chinese. Its points, in summary:
- The attacks had nothing to do with the author.
- ARTEX was built for learning, research and authorized testing.
- The author takes no responsibility for illegal use.
- ARTEX would stop updating and go closed-source, with no public versions and no maintenance.
The statement never mentioned South Korea. By October 10, Reuters reported that the GitHub page was gone.
Korean broadcaster SBS put the problem plainly two days later: software that has already been distributed can't be recalled.
The copies that stayed behind
The developer's exit ended the updates and nothing else.
A week after the statement, Korean security firm AhnLab published an analysis of what remained. It found ARTEX running under new names, including Scope Weaver, Tangora and Horda. Its conclusion: ARTEX is not a single project but is being reused, modified, localized and redistributed by multiple operators.
Other counts tell the same story:
- One code review counted 977 GitHub repositories matching the name on November 2.
- Infosecurity Magazine reported a commercial version called ARTEX Pro.
- Censys counted 1,465 ARTEX control panels exposed to the internet on November 4, and said the true number is likely higher, since a console doesn't have to face the public internet. Some are likely run by authorized testers. Nobody has published how many.
Korea wasn't first
The Korean banks made ARTEX famous, but they weren't the first victims.
- Google's Mandiant incident responders saw attackers operating ARTEX from late August, a month before Shinhan.
- Microsoft saw AI agents flooding GitHub with searches for leaked passwords and keys, also from August.
The biggest case: Microsoft and ConnectWise
Microsoft's October 21 report is the largest documented ARTEX case.
It links a group it calls Storm-0940 to the Chinese state and says the group has been active since 2021. Here's how the campaign unfolded:
- October 10 to 14: ARTEX agents tried to break into 327 accounts at eight organizations. Microsoft logged more than 151,000 failed sign-ins, peaking on October 11 at 2,900 failures a minute. Not one succeeded.
- October 15: the method changed. Storm-0940 took control of the identities of ConnectWise cloud applications, including ConnectWise Cloud Backup, software companies had already authorized to access their Microsoft 365 data. Trusted apps sign in without anyone typing a password.
- Microsoft says the group then used ARTEX to attempt sign-ins to more than 2,250 Microsoft 365 customers.
That last figure counts attempts. Microsoft hasn't said how many customers were compromised and notes that even a successful sign-in doesn't prove data was taken. ConnectWise disabled the affected application identities on October 17.
Microsoft did say that in several ARTEX attacks, the group went from first targeting an organization to privileged access in under three hours.
VAPOR PANDA and the crime groups
By then, others had picked up the tool.
VAPOR PANDA's campaign began on October 17, nine days after Autumn-27 walked away. According to CrowdStrike, the group used ARTEX against more than 200 organizations, including foreign ministries in Latin America, Europe, Asia and the Middle East; US semiconductor, defense and think tank organizations; and telecom providers in several regions. CrowdStrike confirmed follow-on activity at nine of them.
One detail shows the scale of the automation. In a single intrusion, VAPOR PANDA's planning agents consumed roughly 1.4 billion input tokens.
Then on November 6, Microsoft reported a Cambodia-based criminal group, Storm-3358, using ARTEX against financial firms, mostly in Southeast Asia.
Count them: a state-linked group, an espionage group, a crime group and possibly a lone individual, all reaching for the same free software within about ten weeks.
How much did the AI actually do?
This is where the sources split, and where the hype needs checking.
The case for heavy AI involvement: South Korea's Financial Security Institute reportedly estimated that AI, rather than human hackers, led the reconnaissance, intrusion and theft. And the speed is real. Google recorded one burst of 27 events in 23 seconds. Microsoft recorded an agent going from searching GitHub for a leaked key to pulling data out of a company's Azure Key Vault in five minutes.
The case for caution: a Financial Security Institute official also told The Herald Business that the AI did not act independently. A hacker used it as a tool. One of the authors of Google's threat report told iTnews there's no evidence that fully autonomous, human-out-of-the-loop hacking is here yet, describing the campaign as unsophisticated "skids" (script kiddies, or low-skilled attackers) firing into the sky to see what happens, while still calling AI-powered frameworks legitimate.
The tool also makes mistakes a careful human wouldn't:
- Google found ARTEX putting its own name in files and accounts it created.
- Google says these frameworks trade stealth for speed, and it saw long pauses between bursts of activity.
- Check Point found criminal ARTEX campaigns cost just a few hundred dollars to run, and that the agents still hallucinate successes. One write-up of its report says agents collected more than 300 stolen code-signing certificates, and only three passed their own checks.
The Cloud Security Alliance reviewed the Korean evidence and concluded it doesn't establish what share of the operation ARTEX performed on its own.
The number that actually matters
Here's the most useful way to think about it.
Autonomy, the percentage of the hack the AI did alone, may be the wrong number to track. One reviewer who read all 362 Go source files in ARTEX reached a sharper conclusion: ARTEX didn't lower the skill floor for hacking. It lowered the staffing floor for running an attack operation.
A skilled person still has to pick the target and read the results. What that person no longer needs is the team of junior staff who used to run the scans and try passwords through the night.
Microsoft's wording is that agents can work around the clock, on weekends and holidays. CrowdStrike described VAPOR PANDA as one human operator directing multiple simultaneous intrusions. A lone hacker can now work like a whole team, for a few hundred dollars.
As Google's report put it, AI-enabled offensive tooling doesn't fundamentally change attackers' techniques. Stolen passwords and unpatched systems still do the damage. What changes is the clock.
What stopped it
The same evidence shows what worked.
- Multi-factor authentication. Woori Bank and NH NongHyup Bank were targeted and reported no losses. According to Genians' account, South Korea's Financial Supervisory Service concluded that whether MFA was applied is what divided the damage.
- Not trusting forgotten side doors. In Korea, the way in was never the banking app. It was a broker lookup page, an employee work system or a vendor's product with a hole in it.
- Locking down third-party connections. If your IT runs through a provider like ConnectWise, that connection already has access. An attacker who takes it over doesn't need your password.
What you can do
If you write code: Microsoft's November 5 report is aimed at you. An exposed key stays usable until someone revokes or rotates it. Deleting the commit doesn't fix it. GitGuardian counted 28.65 million new hard-coded secrets in public code in 2025. Scan your repos and rotate anything that ever leaked.
If you run IT: enforce MFA everywhere, including internal and partner-facing tools. Audit every third-party app with access to Microsoft 365 or Google Workspace and remove ones you don't need. Review recent changes to credentials, roles and consents.
If you're a customer of an affected bank: watch for targeted phishing using your name, phone number and loan details. In Korea, lawyers began recruiting plaintiffs for class actions on October 7, discussing claims of 100,000 to 300,000 won per person.
For defenders, a clever trick: Microsoft's November 6 report recommends decoys that hand an agent fake wins: convincing evidence that its job is done, so it stops on a false success. Microsoft says these controls cut attack traffic reaching real systems in its own testing. AI's tendency to hallucinate success becomes a weakness you can exploit.
Where this is heading
The rest is projection, and should be read that way.
Google's threat report, dated October 30, 2026, lays out stages of AI-driven attacks. It places today at a stage with fewer humans in the loop during operations, and lists the next stage, where human operators are no longer in the loop, as near-term: one to three months out. If that holds, the first fully autonomous campaigns could appear by the end of January 2027. Google also says with high confidence that criminals will keep adding these tools.
There's a serious reason it might not happen that fast: Google's own analyst says there's no evidence of full autonomy yet, and the agents still report wins they haven't achieved.
Meanwhile, Microsoft's November 6 list includes ARTEX servers it first saw on November 4 and marks them active.
FAQ
What is ARTEX?
An open-source, multi-agent AI penetration testing framework developed in China. A planner agent splits a goal into tasks for worker agents that run standard security tools, using whatever language model the operator connects.
Who has used ARTEX in attacks?
An unidentified, likely Chinese-speaking attacker against South Korean financial firms; Microsoft's Storm-0940 against Microsoft 365 customers via ConnectWise; CrowdStrike's VAPOR PANDA against 200+ organizations; and the crime group Storm-3358.
Can AI hack completely on its own now?
Not according to the evidence so far. Investigators say humans still chose targets and directed operations. The big change is speed and how few people an operation needs.
Did shutting down the GitHub project stop it?
No. Copies, renamed forks and a commercial version kept circulating, and active servers were still being found in November.
What's the single best defense?
Multi-factor authentication. It was the clearest dividing line between Korean banks that lost data and those that didn't.
The bottom line
ARTEX isn't a robot super-hacker. It's a cheap, fast, sometimes sloppy assistant that lets one person do the work of a team, and it can't be put back in the box.
That shifts the defender's job. You can't count on attackers being too busy or too short-staffed anymore. The old basics, MFA, rotated keys, audited vendor access and patched side doors, are what held the line this autumn. They'll have to hold it again, faster.
Sources
- Unknown Threat Actor Uses AI-Driven ARTEX to Target South Korean Finance (CrowdStrike)
- VAPOR PANDA Uses AI-Driven ARTEX to Grow Global Espionage Campaign (CrowdStrike)
- AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise (Microsoft Security)
- Microsoft says ConnectWise breach let Chinese spies loose on 2,250 cloud tenants (The Register)
- Chinese AI tool ARTEX used in wave of bank hacks, probe finds (The Herald Business)
- ARTEX AI Pentesting Tool Used in Data Theft Attacks on South Korean Financial Firms (The Hacker News)
- How AI emerged as new threat in Korea's bank hacking crisis (The Korea Times)





