This defensive guide is based on the FBI’s FLASH advisory as reported by security outlets, plus researcher write-ups. It deliberately leaves out how the tool works in operational detail. Every figure is attributed, and you should read the advisory itself for indicators of compromise.
ARTEX is the clearest sign yet that AI agents have become a mainstream attack tool. On 13 November 2026 the FBI issued a FLASH alert warning that multiple threat actors have used ARTEX, an open-source AI penetration testing framework, since at least August to automate intrusions against organisations in finance, government, technology, defence, healthcare and other sectors.
The headline number is speed. The FBI says some ARTEX-driven intrusions went from initial targeting to privileged access in under three hours, and that the framework lets attackers with minimal technical skill run complex operations against many targets at once. The good news, if there is any, is that the defences the FBI recommends are mostly the basics. The bad news is that the basics now have to be done at machine speed too.
Table of Contents
- What the Tool Is
- What the FBI Said
- Campaigns Linked So Far
- Why This Changes the Threat Model
- Defence 1: Shrink Your External Attack Surface
- Defence 2: Lock Down Cloud Admin Interfaces
- Defence 3: Audit App Permissions and Identities
- Defence 4: Rate-Limit Sensitive APIs
- Defence 5: Watch for Machine-Speed Patterns
- Defence 6: Use the Published ARTEX Detections
- Defence 7: Govern Your Own AI Agents
- A 30-Day Response Plan
- ARTEX Priorities by Team
- FAQ About ARTEX
- Verdict
- Sources and Further Reading

What the Tool Is
ARTEX is an autonomous penetration testing framework written in Go. It was built for authorised security testing, the kind of work red teams do with permission to find weaknesses before criminals do. It coordinates specialised AI agents that handle stages of an assessment with little human involvement.
It has no model of its own. Instead it plugs into external large language models. Reporting on the advisory lists Anthropic’s Claude, OpenAI’s GPT, DeepSeek, GLM, Grok and Kimi among the models it can use. The FBI says analysis of attacker-run instances found Chinese models most common, with DeepSeek’s Flash family appearing most frequently.
A developer using the handle Autumn-27 released ARTEX on GitHub in July 2026. After reports of abuse, the repository was removed on 8 October, with a note that the project would no longer be updated and would become closed source. That did not stop it spreading. The FBI says ARTEX continues to circulate through community forks, localised versions, archived copies and a commercial edition called ARTEX PRO.
What the FBI Said
The advisory, FLASH-20261113-001, was published through the Internet Crime Complaint Center and marked TLP:CLEAR, which means it can be shared without restriction. It runs to 13 pages and includes indicators of compromise.
The key points, as reported by The Record, BleepingComputer and others:
- ARTEX is being used to automate multiple phases of intrusions at machine speed, from reconnaissance through data exfiltration.
- Some intrusions reached privileged access in under three hours.
- Researchers identified more than 1,600 publicly exposed ARTEX consoles as of early November, and the FBI believes real deployments are significantly higher.
- Both criminal groups and nation-state actors are using it.
- It is especially effective against overlooked systems: partner portals, employee applications, legacy services and third-party integrations.
The FBI cautions that the IP addresses in its indicators may already have changed. Treat them as a starting point for hunting, not a complete blocklist.
Campaigns Linked So Far
Several campaigns have been linked to ARTEX or similar AI-agent tooling in public reporting.
| Campaign | What was reported |
|---|---|
| ConnectWise Cloud Backup supply chain (October) | Attackers compromised application identities used by the backup service and authenticated to customer Microsoft 365 tenants. Huntress identified at least 162 compromised tenants; attempts were made against more than 2,250. Microsoft attributed the campaign to the China-state-aligned group Storm-0940. |
| South Korean banks | A Chinese-speaking actor reportedly used ARTEX with Claude Code agents against Shinhan Bank, KB Kookmin Bank and Hana Bank, exposing data on about 67,000 people. |
| South Korean criminal gang | Seven suspects arrested after stealing about 8.2 million records from more than 90 organisations. |
| Vapor Panda (CrowdStrike) | A Chinese espionage group targeted more than 200 organisations in 14 countries and compromised at least nine. |
The pattern across these is not exotic zero-day exploits. It is fast, wide scanning of things organisations forgot they had exposed, followed by abuse of identities and permissions that were broader than they needed to be.
Why This Changes the Threat Model
Penetration testing tools are not new, and attackers have always repurposed them. What makes ARTEX different is the combination of three things.
Speed. Work that took a skilled operator days or weeks now takes hours. Your detection and response times were probably designed around the old pace.
Scale. Because agents run in parallel, one attacker can work against many targets at once. Being a small or obscure organisation is less protection than it used to be.
Skill floor. The FBI’s phrase about minimal technical expertise is the one to remember. ARTEX lowers the barrier to entry. More attackers will be capable of attacks that used to require experience.
The model behind it barely matters. ARTEX works with frontier commercial models and open-weight ones alike, and open-weight models can be run with no usage policy at all. Blaming any one model provider misses the point; the tooling is the threat.
There is an uncomfortable symmetry here. The same agent techniques that let a defender test a network overnight let an attacker do it too, and the line between the two is permission, not technology. Security teams that already use automated testing have an advantage: they know what these tools find first, and they can fix it before someone else looks.
Defence 1: Shrink Your External Attack Surface
The FBI’s first recommendation is to secure internet-facing applications, and OODA Loop’s analysis calls inventorying and shrinking that surface the top priority. If ARTEX is best at finding forgotten portals and legacy services, the most effective defence is to stop having forgotten portals and legacy services.
Start with an honest inventory of everything reachable from the internet: test environments, old marketing sites, partner portals, admin panels and APIs. Anything nobody owns should be switched off or put behind a VPN or identity-aware proxy. This is dull work, and it is the single biggest thing you can do.
Defence 2: Lock Down Cloud Admin Interfaces
The FBI recommends restricting access to cloud administration interfaces. In practice, that means no admin consoles reachable from the open internet without strong authentication, phishing-resistant multi-factor authentication for every admin account, and conditional access rules that block sign-ins from unexpected locations or devices.
If you run self-hosted tools such as automation platforms or agent dashboards, apply the same rule. Our guide to self-hosting n8n covers putting admin interfaces behind authentication and a reverse proxy, and the same principles apply to any web console.
Defence 3: Audit App Permissions and Identities
The ConnectWise case shows why this matters. Attackers did not need user passwords. They compromised application identities that already had permission to access customer Microsoft 365 tenants, and used those permissions as designed.
The FBI recommends reviewing application permissions and authentication logs. List every third-party app with access to your email, files and directory. Remove the ones you no longer use. Reduce the permissions of the rest to the minimum they need. Then check sign-in logs for those app identities, not just for people.
Defence 4: Rate-Limit Sensitive APIs
Automated attacks are, by definition, high volume. The FBI recommends rate limits on sensitive APIs. Login endpoints, password reset flows, search APIs and anything that returns customer data should all have limits that a human would never hit and an agent would.
Rate limits do not stop a determined attacker, but they slow automated tooling down and generate the signals your monitoring needs to spot it.
Defence 5: Watch for Machine-Speed Patterns
Several of the FBI’s recommendations are about monitoring: unusual bursts of automated requests, suspicious cloud activity, unexpected credential or privilege changes and large data transfers.
The common thread is tempo. A human attacker pauses, explores and makes mistakes over days. An ARTEX-driven intrusion can chain steps in minutes. Alert on sequences that happen too fast for a person, such as a new admin role granted moments after a first login from a new location, followed by a large export. And make sure someone is watching those alerts outside office hours, because a three-hour attack can start and finish overnight.
Defence 6: Use the Published ARTEX Detections
You do not have to build detections from scratch. The FBI’s advisory includes indicators of compromise, and on the same day the research group MalwareMustDie released an open-source ARTEX detection signature pack covering network requests, identity abuse, cloud activity, malicious files and automated attack patterns.
That playbook sorts indicators by confidence, because infrastructure indicators go stale quickly and can produce false positives. Load the high-confidence behavioural detections first, and treat IP-based indicators as hunting leads. If you find anything, the FBI asks for reports to its field offices or IC3.
Defence 7: Govern Your Own AI Agents
The last defence is closer to home. ARTEX started life as a legitimate tool, and the same agent techniques power the assistants your own teams are adopting. The South Korean bank case reportedly involved Claude Code agents directed by the attacker.
Know which AI agents run inside your organisation, which credentials they hold and what they can reach. Give them their own scoped identities rather than personal logins, log their actions and review their permissions as you would for any service account. Our guides to agent security and AI agent frameworks cover how to build those controls in.
If your security team uses autonomous testing tools legitimately, keep their consoles off the public internet. More than 1,600 exposed ARTEX consoles were found online; some of those may belong to defenders who did not lock them down.
A 30-Day Response Plan
Advisories are easy to read and hard to act on. Here is how I would turn this one into a month of work for a small or mid-sized organisation, without hiring anyone new.
Week one: see what is exposed. Run an external scan of your own domains and IP ranges, or use your cloud providers’ built-in tools to list public endpoints. Compare the result with what you think you have. The gap between the two lists is your real risk. Assign an owner to every item or switch it off.
Week two: fix identity. Turn on phishing-resistant multi-factor authentication for every administrator. Review third-party applications with access to email, files and directory, and remove or reduce anything you cannot justify. Check that service accounts and app identities have expiry dates and owners.
Week three: add speed bumps and sensors. Put rate limits on login, password reset and data export endpoints. Turn on alerts for privilege changes, new admin roles and unusually large downloads. Load the advisory’s indicators and the open-source detection pack into whatever logging you already have, even if that is just your cloud provider’s native tools.
Week four: rehearse. Pick one scenario, such as a compromised partner account reaching your email tenant, and walk through it as a team. Who gets the alert? Who can revoke access at 2am? How fast can you rotate an app’s credentials? The answers will show you where the next month’s work should go.
None of this requires new tools. It requires attention, and the willingness to switch off things that people have quietly relied on for years. That is uncomfortable, and it is exactly the work automated attackers are counting on you to postpone.
Finally, write down what you changed and when. If you ever need to report an incident, or prove to a customer that you acted on a public warning, that record will matter.
ARTEX Priorities by Team
| If you are | Start with | Why | Link |
|---|---|---|---|
| A small business with no security team | MFA everywhere and an inventory of what is online | Attackers target forgotten, exposed systems first | FBI IC3 |
| A Microsoft 365 admin | Review third-party app permissions and sign-in logs | The ConnectWise campaign abused app identities | BleepingComputer report |
| A security operations team | Load the FBI IOCs and MalwareMustDie detections | Ready-made detections, sorted by confidence | The Record’s coverage |
| A team deploying AI agents | Scoped identities and action logs for every agent | The same techniques power legitimate and malicious agents | Agent security guide |
| A red team using autonomous tools | Take consoles off the public internet | Over 1,600 exposed ARTEX consoles were found | OODA Loop analysis |
Related on this site: for more context, see Mistral Large 4.
FAQ About ARTEX
What is ARTEX?
ARTEX is an open-source AI-driven penetration testing framework written in Go. It coordinates AI agents that automate stages of a security assessment, and the FBI says attackers have repurposed it to automate real intrusions.
What did the FBI warn about ARTEX?
In FLASH-20261113-001, issued on 13 November 2026, the FBI said multiple threat actors have used ARTEX since at least August, with some intrusions reaching privileged access in under three hours.
Which AI models does ARTEX use?
It has no model of its own and connects to external ones. Reports list Claude, GPT, DeepSeek, GLM, Grok and Kimi, and the FBI says Chinese models, especially DeepSeek Flash, are most common in attacker deployments.
Is ARTEX still available?
The original repository was removed on 8 October 2026, but the FBI says forks, archived copies, localised versions and a commercial edition called ARTEX PRO remain in circulation.
How do I protect my organisation from ARTEX?
Follow the FBI’s recommendations: secure internet-facing apps, restrict cloud admin access, review app permissions and logs, rate-limit sensitive APIs and monitor for automated bursts, privilege changes and large transfers.
Where can I find ARTEX indicators of compromise?
In the FBI’s FLASH advisory, published through IC3, and in the open-source detection pack MalwareMustDie released the same day. The FBI notes that IP indicators may have changed.
Verdict
Take the ARTEX warning seriously, and respond with fundamentals rather than panic. The tool is fast, widely available and easy to use, but it wins mostly by finding forgotten systems and overly broad permissions. Shrinking your attack surface and tightening identities blunts most of what it does.
The caveat is that removing the original repository changed nothing, and similar frameworks will follow. ARTEX is a sign of where attacks are heading, not a one-off, and detection has to move closer to machine speed to keep up.
Your next step: this week, list every internet-facing system and every third-party app with access to your email and files. Switch off what nobody owns, cut permissions on the rest, and load the FBI’s indicators into whatever monitoring you have.
Sources and Further Reading
- The Record on the FBI flash alert, for the advisory’s key findings and quotes
- BleepingComputer, for the ConnectWise supply-chain breach and sectors targeted
- OODA Loop’s analysis, for the advisory format and the top recommendation
- Petronella Technology Group, for the full mitigation list and related campaigns
- FBI Internet Crime Complaint Center, where the advisory is published and incidents can be reported
More on agent security on this site: prompt injection and agent security, AI agent frameworks compared and building an agent from scratch with guardrails in place.

