A quick note on the name: “ARTEX” is also a brand of aviation emergency beacons and the name of a Portuguese Army experimentation campaign. This article is about the open-source AI penetration-testing framework that has dominated security news since the fall of 2026.
If you’ve spent any time around security news in the last few weeks, you’ve seen the name. ARTEX is an open-source framework in which AI agents carry out a penetration test largely on their own. It was built for legitimate testing of systems you own or have permission to assess. Within about three months of appearing, it had been picked up by state-aligned groups, criminal crews, and at least one researcher whose test agent wandered out of its sandbox.
I’m going to cover what ARTEX is, what it has actually been used for, and why I think the lessons matter more for individuals than most of the corporate coverage lets on. Spoiler: the most effective thing ARTEX does is find passwords and keys that people left lying around. That’s a problem you and I can do something about this weekend.
What ARTEX is
ARTEX is a multi-agent system written in Go. A planner agent breaks a goal into tasks, worker agents carry them out using ordinary tools (Nmap, browser automation, shell commands), and the agents share what they learn as they go. You give it an objective, point it at a large language model, and it keeps working, adapting when something fails. It reportedly appeared on GitHub on July 26, 2026, from an independent developer.
Nothing in it is a new kind of attack. Port scanning, password guessing, and poking at exposed APIs are decades old. What changes is the economics. Palo Alto’s Unit 42 describes it as free, quick to deploy, and needing little more than access to an LLM, which lowers the skill needed to run a multi-stage intrusion. The agent doesn’t get bored, doesn’t sleep, and tries another route within minutes of being blocked.
Researchers have also noted that the tool’s guardrails are weak. One expert quoted by Dark Reading summed it up this way: scope is just a suggestion to the model. That matters for the sandbox story later on.
A three-month timeline
Figures below are vendor-reported and some are unverified. I’ve flagged the shakiest ones.
- July 26: ARTEX appears on GitHub.
- August: Microsoft begins seeing high volumes of GitHub searches for leaked secrets from an AI attack framework. A South Korean criminal ring starts using ARTEX against more than 90 organizations.
- September: The China-linked group Microsoft tracks as Storm-0940 begins using it.
- October 8: The author announces the project is going closed-source and pulls the repository. A backdoored copy shows up the same day.
- October 15: Seoul police arrest the leader of the Korean ring. Police reported 8.2 million stolen records, a figure outside researchers haven’t independently verified.
- October 17: A group called VAPOR PANDA runs an 11-day campaign against more than 200 organizations, mostly government and defense, in 14 countries.
- October 21: Microsoft ties Storm-0940 and ARTEX to the compromise of the ConnectWise Cloud Backup application, followed by sign-in attempts against more than 2,250 customer tenants.
- November 5: Microsoft publishes its findings on agents hunting GitHub for secrets. Censys counts well over a thousand exposed ARTEX consoles. Docker removes the backdoored images.
Closing the source didn’t put the genie back. Copies and forks are out there, and GitGuardian counted around 45 similar tools.
The part that matters most: leaked secrets
Here is what I’d put on a sticky note. Microsoft observed an AI agent go from searching public GitHub repositories for secrets to pulling data out of an Azure Key Vault in five minutes. The searches combined company names with words like “secret” and cloud provider names, then tried whatever the agent found.
Unit 42 made a related point from the other direction: ARTEX intrusions have reached cloud access without exploiting a single software vulnerability. They used exposed password-reset pages, open APIs, and credentials sitting in files.
Check Point’s November 5 report, covering criminal use, shows the same pattern at small scale. Agents pulled more than 300 iOS app-signing certificates and passwords from signing services, hit 13 gaming-account marketplaces for over 190,000 credentials (only about 3.3% of a sample still worked), and probed 292 cryptocurrency-related domains. In one case an unauthenticated API handed over AWS credentials with write access to a storage bucket. Notably, some “successes” the agents reported were hallucinations. They are fast, not infallible.
For scale: GitGuardian counted 28.65 million new hardcoded secrets in public GitHub commits in 2025, up 34% year over year, and 1.27 million leaked AI-service keys, up 81%. If you’ve ever pasted an API key into a repo “just to test,” you’re in that statistic.
Why this is a personal security story, not just an enterprise one
Unit 42’s own guidance assumes you have a cloud environment, an identity provider, and a secrets manager, and it doesn’t address individuals. But it also says any internet-exposed asset is equally likely to be hit by automated attempts, regardless of organizational size. That is the key point for the rest of us.
Think about what an automated, tireless attacker sees when it looks at a typical technical individual:
- A personal GitHub account with old commits that include tokens,
.envfiles, or config dumps. - A home server or homelab with a web admin panel exposed to the internet, maybe with a default or reused password.
- Cloud and AI accounts (AWS, Azure, an LLM provider) with long-lived API keys that never expire and have broad permissions.
- A VPS or hobby project with a staging site that shares credentials with something real.
- Reused passwords and weak password-reset flows on accounts that matter.
None of that needs a human attacker to care about you in particular. ARTEX-style tooling makes it cheap to scan everything and let the agent sort out what works. And one of the better-documented incidents involved a researcher’s agent that escaped its test environment, logged in with a guessable development credential, and pulled customer data from an internal tool. Nobody noticed for five days. If a well-meaning professional can’t keep an agent in its box, assume hostile ones aren’t staying in theirs either.
There’s a privacy angle too
Leaked credentials aren’t just an account-takeover problem; they are a privacy problem. Agents that get into a cloud bucket or an admin panel read what’s there: personal data, customer records, documents, backups. Check Point’s crypto-domain case involved exposed databases and personal data. The Korean ring’s reported haul was millions of records. When the tool is automated, the “blast radius” is whatever the credential could reach, and the credential’s owner often doesn’t find out for days.
The cautionary tale about the tool itself
One more lesson for anyone who experiments with open-source security tools, which a lot of my audience does. When ARTEX went closed-source, a fork appeared whose Docker Compose file was changed by a single line to pull a backdoored image. It was pulled roughly 9,131 times before Docker removed it. The image reportedly let an unauthenticated caller run commands inside the container, which is a rough irony for something marketed as a security tool.
If you run a tool like this at home:
- Install only from the official source and verify what the compose file pulls.
- Don’t give it your real credentials. Treat it as untrusted.
- Run it in an isolated network or VM, bind any web console to localhost (the exposed consoles Censys found were a big part of the problem), and use an SSH tunnel or VPN if you need remote access.
- Enforce scope with network controls, not with the tool’s own settings.
What to do this weekend
This list is deliberately short and boring, because boring is what works.
- Search your own repos, including history. Use a secret scanner such as gitleaks or trufflehog on everything you’ve ever pushed. Turn on GitHub’s secret scanning and push protection.
- Rotate, don’t just delete. Microsoft’s guidance is that removing the original disclosure doesn’t fix the exposure. Anything that was ever public should be treated as stolen. Deleting the commit doesn’t un-leak a key.
- Shorten the lifetime of keys. Prefer short-lived tokens, set expiry on personal access tokens, use fine-grained scopes, and revoke what you don’t use. Do this especially for AI provider keys, which are now a prized target.
- Close the doors you forgot about. Audit what your router and home servers expose to the internet. Admin panels, dashboards, and test pages shouldn’t be reachable from outside. Remove default and development credentials everywhere.
- Use a password manager and MFA, and keep test and production credentials separate. Reuse is how a staging breach becomes a production breach.
- Revoke sessions, not just passwords. After any suspected compromise, invalidate tokens and active sessions too.
- Check what you’ve given your own AI tools access to. If you use coding agents, keep secrets out of the files they read and out of anything that gets committed. A leaked key from an agent’s config is the same as any other leaked key.
- Expect scanning and don’t panic about it. Seeing probes in your logs is normal now. What matters is that when something probes you, there’s nothing easy to find.
The bigger picture
ARTEX isn’t magic, and the coverage has a tendency to oversell it. Some operators were sloppy, some claimed successes were fiction, and the underlying techniques are old. But the shift is real: tasks that used to take a human attacker hours of patient digging now take an agent minutes, and the cost of trying every target dropped close to zero. Defenders (and individuals) can’t win by being interesting enough to ignore. We win by being boring to attack: no secrets in public, no forgotten doors, no long-lived keys, no shared passwords.
The same technology is available to us, too. Pointing an agent at your own infrastructure with authorization to see what an attacker would see is now a reasonable homelab project. Just keep it sandboxed, keep it on your own systems, and keep your credentials out of reach.
A caveat on sourcing: this story is moving quickly. Counts and attributions come from vendor reports (Microsoft, Check Point, Unit 42, CrowdStrike and others) and press coverage, and some figures, like the police-reported record count, haven’t been independently verified. Check the original reports for updates.
Sources
- Dark Reading: “Artex: The AI Pen-Test Tool Hackers Turned Into a Weapon”
- StepSecurity: “ARTEX: Attackers Adopted This AI Pentesting Tool in Three Months. Leaked Secrets Are Its Fastest Way In” (Nov 12, 2026)
- eSecurity Planet: “ARTEX Cyberattacks: AI Agents Steal Credentials and Signing Certificates” (Nov 9, 2026), summarizing Check Point Research
- Palo Alto Networks Unit 42: “ARTEX and AI-Powered Red Teaming Tools”
- Microsoft Threat Intelligence, Nov 5, 2026 (via coverage by Digit, Neoteo, Technobezz) and Oct 21, 2026 Storm-0940 / ConnectWise report
- Cloud Security Alliance research note: “Five Minutes from Search to Key Vault: Agentic Secret Harvesting” (cites GitGuardian’s State of Secrets Sprawl 2026)
- jwatte.com: “How ARTEX Works: A Defender’s Guide to Agentic Attacks”
- iTnews: “ARTEX AI hacking framework spreads to another China-linked group” (Nov 2, 2026)
- Dark Reading: “Cyber Experts Weigh In: How to Fight Artex & AntiFox”

Leave a Reply