ARTEX detection is harder than it was a month ago, because the framework no longer travels under one name. Our ARTEX profile explains what it is: an open-source AI penetration testing framework that attackers have used since late August 2026, according to Google Threat Intelligence Group (GTIG). This playbook covers how to find it and what to do first.
What Does the FBI Flash Alert Say About ARTEX?
The FBI’s alert, FLASH-20261113-001, says a variety of malicious actors have used ARTEX since at least August 2026. Targets include organizations in financial services, government, technology, healthcare and telecommunications. It says some intrusions went from initial targeting to privileged access in under three hours. It also says researchers identified over 1,600 publicly exposed ARTEX consoles in early November.
The alert carries indicators, which the FBI says to vet before blocking because cloud addresses get reassigned. Its mitigations, in short:
- Inventory every internet-facing application, including partner portals, sales-support systems and legacy services.
- Restrict cloud administration and Kudu or SCM publishing endpoints to trusted networks.
- Apply least privilege to application permissions and consent grants, and review service-principal sign-ins and credential additions.
- Add rate limits to business APIs, and alert on sequential lookups and unusually many records per session.
- Scan code, images and pipelines for embedded secrets, and rotate any that are exposed.
- Preserve Entra sign-in and audit records before cleanup.
Why Does Name-Based ARTEX Detection Miss Forks?
The FBI says the original developer made the ARTEX repository unavailable on 8 October 2026. Wiz reports the code already had more than 12,000 forks by then. The FBI lists forks, localized variations, archived copies and a commercial version called ARTEX PRO as the ways it keeps spreading. A hunt for the name ARTEX therefore misses a renamed copy, so in our view a threat intelligence platform has to follow attacker tools by behavior rather than by name.
Talos describes ScopeWeaver as a localized derivative, rebranded and translated into English and Korean, and AntiFox as a delivery-enhanced version of ARTEX. Arctic Wolf tied Alibaba Cloud hosts running AntiFox consoles to confirmed intrusions on 10 November 2026. Censys saw consoles branded ARTEX MX among other renamed builds. The fork passage on the ARTEX profile names each one.
Brandefense Blackhole recorded 25 posts naming AntiFox between 30 October and 14 November 2026, 13 of them on 11 November. Most were feed channels resharing coverage, so read the count as attention, not as use.
Which ARTEX Fingerprints Survive a Rename?
Some traits come from the code, not the branding. These are the ones primary sources publish for detection, and none is proof on its own.
| Fingerprint | Who published it | How far to trust it |
|---|---|---|
| Console page title starting with ARTEX, plus Next.js assets | Censys | Covers localized versions. A fork that drops the name from the title escapes it. |
| Default console port 8787 | Censys, Arctic Wolf | Default for ARTEX. Arctic Wolf lists it as the common AntiFox port. Censys also sees 8790, 443 and 80. |
| AntiFox console title | Arctic Wolf | Specific to one fork. The advisory publishes the title for matching. |
| User agent norma/0.4 | Wiz | Wiz lists it among indicators for an ARTEX-using actor. Talos says user agents can change. |
| Snort rules for norma artifacts | Talos | Talos says they detect artifacts of norma, the agent framework used in ARTEX. They are policy rules, off by default. |
| Outbound calls to LLM provider API endpoints | FBI | The framework needs an external model. Investigate unexpected calls together with other signs. |
Talos says tool signatures and user agents can change, so treat them as supporting indicators rather than attribution. The FBI asks readers to corroborate its indicators with current telemetry. Our reading is to use fingerprints to find candidates, then confirm with behavior.
Which Behaviors Give ARTEX Away Under Any Name?
GTIG measured ARTEX execution sequences in bursts of roughly 2 to 25 seconds, including 27 events in 23 seconds in one operation. Alert on sudden runs of API calls, permission tests and resource listing from one identity. Our post on detection telemetry for five agentic attack frameworks shows how to join those events to the identity behind them.
Arctic Wolf describes a pattern from AntiFox-linked infrastructure. In one intrusion the source cycled through several AWS identities and sent hundreds of cross-service requests in about half an hour. Most were denied, but a successful GetSecretValue request against a production secret sat inside the same burst. Arctic Wolf says not to assume a sequence failed because most requests were denied.
Our post on agentic AI and exposed secrets covers how agents use leaked keys. The same logic applies to any identity that suddenly lists many unrelated services.
Microsoft’s Storm-0940 report shows the identity side. After a vendor’s application credentials were compromised, ARTEX attempted sign-ins to more than 2,250 tenants. Microsoft lists Defender XDR alerts such as an application assigning itself the Global Administrator role.
The FBI adds credential changes on existing OAuth applications and newly created accounts with naming patterns. Related cases: Storm-0940 against ConnectWise customers, Storm-3358 against financial organizations and ARTEX agents hunting AI provider keys.
What Must a Red Team Check Before Running Any ARTEX Copy?
Badbyte reports that a prebuilt ARTEX Docker image carried a backdoored login handler. It ran any command sent to it before checking the password, and the GitHub source matched upstream, so a code review found nothing. Docker removed the image on 5 November 2026 at Badbyte’s request, but Badbyte says copies already pulled stay backdoored.
- Record where the copy came from. With the upstream repository gone since 8 October, a copy found by search may be a fork, a rebrand or a tampered build.
- Build from source you have reviewed instead of pulling a prebuilt image. Badbyte says the backdoor lived only in the image.
- Check what the compose file exposes. Badbyte says the one it reviewed published the console port on every interface and ran the container as root.
- If a suspect image ran, treat model keys, database passwords and harvested credentials as exposed. Badbyte says ARTEX stores such secrets in plaintext.
ARTEX Detection Playbook: What to Do First
The FBI, Arctic Wolf and Talos guidance sits behind these steps. The order is ours.
- Inventory exposureList internet-facing applications, including partner portals and sales tools, plus cloud administration and Kudu or SCM endpoints. Restrict them to trusted networks, as the FBI advises.
- Keep the logsPreserve Entra sign-in and audit records, CloudTrail management events, S3 data events for sensitive buckets and web server logs before any cleanup.
- Hunt bursts by identityIn CloudTrail, group calls by identity and look inside bursts of denied calls for successful secret reads, such as GetSecretValue after ListSecrets. Also look for one source address using several keys.
- Hunt fingerprintsSearch web and proxy logs for the norma/0.4 user agent. Scan your own address space for consoles with an ARTEX-style title on port 8787.
- Check web serversAlert on IIS worker processes that spawn command shells, and on new script files in upload folders.
- Contain and rotateReview each affected identity’s CloudTrail history, then revoke its sessions and access keys. Rotate secrets stored in application configuration.
Exposure and logs come first because every later hunt depends on them, and the FBI says attack speed leaves defenders little time to react. Fingerprints come after behavior. In our reading, a renamed fork can change a title or user agent more easily than its burst pattern or identity trail.
To check whether AI hacking tools such as ARTEX have targeted your own organization, Atlas, Brandefense’s free security intelligence platform, offers an AI Hacking Tools Check.
Frequently Asked Questions
How do you detect ARTEX when a fork has renamed it?
Combine fingerprints with behavior. Censys publishes a console title and port pattern, Wiz lists the norma/0.4 user agent, and Arctic Wolf describes bursts of denied cloud API calls with successful ones among them. Talos says to treat tool signatures as supporting indicators.
Does the FBI FLASH list ARTEX indicators?
Yes. FLASH-20261113-001 of 13 November 2026 lists addresses, a domain and network and behavioral indicators. The FBI says to vet them before blocking, because cloud addresses can be reassigned.
Can an ARTEX Docker image carry a backdoor?
Badbyte reports that one prebuilt image did, through a modified login handler. It says the GitHub source matched upstream, and that Docker removed the image on 5 November 2026.
What should a red team check before running an ARTEX copy?
Check where the copy came from and build from reviewed source. Keep the console off the internet. If a suspect image ran, treat the keys and passwords it stored as exposed.
Sources
- FBI FLASH-20261113-001, Threat Overview of ARTEX Autonomous Penetration Testing Framework, 13 November 2026
- Cisco Talos, ARTEX and beyond: AI-enabled penetration testing platforms, 13 November 2026
- Arctic Wolf Labs, Antifox agentic pentesting platform advisory, 10 November 2026
- Badbyte, ARTEX Backdoor, 4 November 2026 (updated 5 November)
- Censys, ARTEX: Agentic Asbestos, 5 November 2026
- Wiz Research, ARTEX in the Wild: Storm-0940 Expands Targeting of IT and Software Vendors, 22 October 2026
- Google Threat Intelligence Group, ARTEX and Friends: Enter Agentic Offensive Frameworks, 30 October 2026
- Microsoft Threat Intelligence, AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise, 21 October 2026
