ARTEX: The Pentest Tool That Became Everyone's Attack Team
The FBI says threat actors have used ARTEX, an open-source AI pentest framework, since August to automate intrusions, some reaching privileged access in under three hours. Where it came from, how it works, the campaigns behind it, and how to detect and stop it.
ARTEX was built to help defenders. A developer published it on GitHub in July as an autonomous penetration-testing tool: you describe a goal in plain English, and AI agents do the testing for you. Within ten weeks it had an estimated 1,500 to 2,000 stars. Within a month of release, attackers were using it too.
On 13 November the FBI published a FLASH alert, FLASH-20261113-001, saying several threat groups have used ARTEX since at least August to automate intrusions from reconnaissance all the way to data theft. Some went from first look to privileged access in under three hours. The targets include banks, government, healthcare, defence, energy and telecoms.
I want to take this one apart properly, because ARTEX is not just another malware family. It is a preview of how attacks will be run from now on: one person, a chat box, and a team of AI agents doing the work. Here is where it came from, how it works, what it has already done, and how to spot and stop it.
It took about a month for a pentest tool to become a weapon
The timeline tells the whole problem in one line. The tool was released in July, abused by August, and taken down by its own author on 8 October. By then it did not matter. The FBI says it lives on through community forks, archived copies, translated versions and a commercial edition called ARTEX PRO, and the advisory says it picked up an estimated 1,500 to 2,000 GitHub stars within 10 weeks.
This is the uncomfortable truth about open-source offensive tools in the age of AI: once a capable one is published, you cannot unpublish it. Defenders have to assume it is in the wild for good.
A chat box on top, a team of agents underneath
The design is simple, and that is what makes it dangerous. The operator types a goal, for example “get admin access to this company’s cloud”. ARTEX breaks it into tasks and hands each one to a specialised agent: one maps the target, one looks for weaknesses, one tries exploits, one climbs to higher privileges, and one takes the data. When an attempt fails, the agents try another path. One operator can run this against several organisations at the same time.
Notice the box at the bottom: ARTEX has no brain of its own. It rents one from commercial or open AI models, and swapping models does not change its behaviour. That is why I would not build defences around which AI the attacker uses. Build them around what the attack does.
Four campaigns the FBI tied to ARTEX
ConnectWise Cloud Backup
Attackers took over the backup service’s application identities and used permissions customers had already granted it. Microsoft tied it to Storm-0940; sign-ins were tried against 2,250+ Microsoft 365 tenants and Huntress confirmed at least 162 compromised.
Three major banks
A Chinese-speaking actor combined ARTEX with Claude Code agents against Shinhan, KB Kookmin and Hana Bank. About 67,000 people’s personal data was exposed.
90+ organisations
Police arrested seven suspects. ARTEX was used in most of their intrusions, and about 8.2 million records were stolen from hospitals, schools and businesses.
Vapor Panda
A Chinese espionage group pointed ARTEX at 200+ organisations. At least nine were compromised; the agents deployed web shells and switched paths when blocked.
Two things stand out across these four. First, the range of attackers: a state-aligned espionage group, criminals chasing bank data and a gang caught by Korean police all used the same free tool. ARTEX lowers the skill needed to run a serious intrusion.
Second, and this is the hopeful part, most of these attacks did not need new zero-days. Reporting on the campaigns points to exposed services, weak credentials and excessive permissions. ARTEX is a speed multiplier for finding the gaps we already have. Close the gaps, and the speed matters much less.
ARTEX is fast, but it is not quiet
Behaviour worth alerting on
- !Machine-speed bursts: Google’s threat team saw ARTEX work in bursts of about 2 to 25 seconds.
- !Rapid sequential requests against login and search endpoints.
- !AI model API calls from systems with no business reason to make them.
- !New .aspx or .jsp files in web paths where they should not be: a sign of web shells.
- !Sudden permission changes by cloud apps or service principals.
What not to rely on
- ?IP blocklists alone. The FBI warns its IP indicators may have changed and overlap with shared cloud infrastructure.
- ?File hashes. The FLASH contains none; it is a framework, not a single binary.
- ?Blocking one AI provider. Attackers simply switch models.
- ✓Better: behaviour-based detections. A free, community signature pack by MalwareMustDie is a good starting point.
Here is the twist. The FBI notes that ARTEX traffic is noisy enough to trigger alerts. The problem is not that we cannot see it; it is that by the time a human reads the alert, the agents may already be done. So detection has to be paired with fast, automatic containment for the clearest signals.
Could you see an ARTEX intrusion today?
Tick the signals your team can see and act on right now. The result shows your coverage and what is missing.
Be strict with yourself: “we have the logs somewhere” does not count. Tick a box only if an alert would reach a person, or trigger an action, within minutes.
What I would put on the board on Monday
These come straight from the FBI’s mitigations and the lessons of the ConnectWise campaign, in the order I would do them.
Clean up app permissions for good
Every third-party app with access to your Microsoft 365 or cloud tenant is a door. Review them quarterly and remove what is not needed.
Automate the first response
For high-confidence alerts, isolate the host or disable the identity automatically. Machine speed needs a machine-speed answer.
Test yourself with the same tools
Run authorised AI-assisted testing against your own estate, so you find the gaps before someone else’s agents do.
The attacker’s team now fits in a chat window.
ARTEX did not invent new attacks. It took the ones we already know and made them cheap, fast and available to almost anyone. The answer is not panic, and it is not a new silver-bullet product. It is doing the basics faster than an AI agent can find the gaps: fewer exposed services, stronger credentials, tighter permissions, and a response that does not wait for a meeting.
Related on DKCISSP: AI agents found a zero-day in minutes · When cyberattacks start thinking
Sources
- BleepingComputer: FBI warns ARTEX AI hacking tool is powering global cyberattacks
- Becker’s: ARTEX targets healthcare and nine other sectors
- DamSecure: how a pentest tool ended up in real attacks
- Brandefense: ARTEX profile and detection notes
- SafeBreach: FLASH-20261113-001 behavioural indicators
- Mallory: attacks rely on exposed services and weak credentials
- BleepingComputer: ConnectWise Cloud Backup supply-chain attack
- Emry Networks: Storm-0940 and 2,250+ tenants
- Tidal Cyber: ARTEX looked for the gaps first
Facts as reported up to 21 November 2026, based on public reporting of FBI FLASH-20261113-001 and vendor research.
- AI Security
- Threat Intelligence
- ARTEX
- Opinion