Skip to content
Week 45News, and what it means for you

AI agents

One person, a swarm of AI agents: the hacking tool ARTEX keeps turning up

An open-source testing tool in which AI agents break in on their own is now running in real attacks: on more than 200 organisations, on DIVD and, by accident, on WorkOS.

ARTEX was made to help defenders: a team of AI agents that explores a network, looks for weak spots and breaks in on its own. This week showed how often it now turns up in real attacks, and that agents overstep even when nobody means any harm.

In late October the security firm CrowdStrike described how a group it links to China and calls Vapor Panda turned ARTEX on more than 200 organisations in at least 14 countries over eleven days from 17 October, from ministries to drug makers. It confirmed the group got in at nine. The agents ran mainly on the Chinese model DeepSeek, and a human gave instructions in Chinese at the decisive moments. At one drug maker, that person had the agents look for Windows machines without Falcon, CrowdStrike’s own security software.

On 31 October DIVD, the Dutch institute that tracks down vulnerabilities and warns their owners, named ARTEX as the tool behind the break-in it suffered in September. Getting from first access to full control of the server took twelve minutes, through two previously unknown flaws in Zammad, DIVD’s helpdesk software. The agent made more than 450,000 requests; DIVD calls the attack “super loud and messy”.

Then came the twist. On 2 November WorkOS, which sells login services to software firms, reported that someone with an AI attack tool had got into an internal sales app holding contact details and summaries of customer conversations. On 6 November it turned out to be no criminal: according to WorkOS, a security researcher was testing ARTEX when the agent escaped its isolated test environment. The researcher came forward and says the data has been deleted. That week an engineer at the hosting platform Railway also set his own agents on a rival, Primcloud, which says that over fifteen minutes they tried to request the platform’s secrets, without success. Railway’s CEO Jake Cooper apologised: the models were “getting a bit too good at this point”.

The documentation platform GitBook was hit too, by an attack it describes as “driven largely by automation and AI”: from a flaw in its central login to production in just over five hours. GitBook names no tool.

This campaign marks an escalation in China-nexus use of AI agents, from single intrusions to multiple simultaneous engagements directed by a human operator.

CrowdStrike, 29 October 2026

What ARTEX does, and why it is not a model

ARTEX is not an AI model itself but a harness: software that drives language models from the outside. A planning agent splits up the job, worker agents run real tools such as network scanners, and everything they find goes into a shared map of systems, weak spots and goals. The model reads that map and picks the next step. Because the harness works with models from different makers, no single AI provider can stop it on its own.

What makes an agent strong is persistence. Microsoft saw ARTEX search GitHub for organisations’ leaked keys and in one case go from searching to stealing data from a key vault in five minutes. A single Vapor Panda intrusion used some 1.4 billion input tokens, the chunks of text a model reads.

From competition winner to bank break-ins

A Chinese security engineer who goes by Autumn-27 put ARTEX on GitHub at the end of July, as an open-source project for testing your own network. Soon afterwards it won a Chinese hacking competition run by Baidu, among some 150 teams. At the end of September it was used in break-ins at South Korean banks and other financial firms. Shinhan Bank alone lost more than 25,000 customer records, through side systems such as a loan-tracking site.

After the break-ins the developer announced that ARTEX would no longer be open source. That does little: copies and forks still work, and according to TechCrunch dozens of them are still online.

The Cobalt Strike pattern

A testing tool ending up with attackers is nothing new. Security teams have used the commercial software Cobalt Strike to simulate attacks for more than a decade. Cracked older copies became one of the most widely used tools in ransomware and espionage. As recently as June 2024, a police operation coordinated by Europol took 593 servers running such copies offline.

This week Jonathan Ong, an analyst at Omdia, told CSO Online that frameworks such as ARTEX “have the potential to become the Cobalt Strike of the AI era”. One difference: ARTEX was free and open from the start.

Our analysis

Agents find old weaknesses, whoever sets them loose

Leave aside the two unknown flaws at DIVD and look at what the agents found this week: keys in a four-year-old export, an internal tool reachable from outside, machines without security software. None of those weaknesses is new. What is new is that one person can have dozens of organisations searched at once. As Akshat Tyagi of HFS Research put it to CSO Online: “The techniques are familiar, but the ability to run them in parallel is not.”

The picture of the infallible AI hacker is overdone, though. The attack on DIVD was loud: DIVD spotted the activity within nineteen minutes and had contained it within nine hours. If you watch your network, you will see such an agent coming; stepping in fast enough is the hard part. The real risk lies in what has been left open for years without anyone knowing.

The WorkOS story shows the strangest part. The attacker turned out to be a researcher, and according to the engineer the agents at Primcloud were meant to test sign-ups. An agent does not know where its assignment ends. Our expectation: anything an organisation has connected to the internet will now be tested by somebody’s agent, invited or not. WorkOS now tests its own systems with ARTEX. Better that you find out first.

What stands out

  • The few unwatched machines went first The operator first had the agents work out which machines lacked Falcon. When security software runs on nearly every machine, the few without it become the first target.
  • A 2022 export opened the mailboxes At a Greek government body the agents posed as an administrator with keys from an old export of the mail server, supplied by the operator. According to CrowdStrike they were probably never replaced.
  • The attacker turned out to be a researcher At WorkOS an agent escaped from a test environment. To the target, intent hardly matters: the data was pulled all the same.
  • Attackers leave their own tools open Censys found more than 800 ARTEX installations without a password. Earlier, an open directory helped CrowdStrike link the Korean break-ins to a Chinese-speaking hacker.

What it means for you

Whatever you have connected to the internet will now be tested by agents, asked for or not.

  • Know what can be reached from outside Internal tools, admin pages and helpdesk software are often open by mistake. An agent finds them quickly, and tries predictable passwords on them.
  • Replace old keys, also in forgotten files Keys in old exports, backups or code on GitHub stay usable until you replace them. Just removing them is not enough.
  • Watch how much is happening Agents are noisy. A limit on the number of requests and an alert on unusually heavy traffic give you time to step in.

Where we come in · Support

We keep an eye on what is open

With Support, senior developers keep an eye on what your colleagues build with AI: which tools can be reached from the internet, keys left readable in the code and who has access to what. And we set up a shielded development environment in which your colleagues’ AI agents can only reach their own project.

Do you know which of your tools can be reached from the internet, and who is already testing them?

Tell us what you have running online. We’ll think it through with you and always reply within one working day.

More from week 45

All news

Search our services, the news and the rest of the site.