The FBI’s ARTEX Warning Says a Supply Chain Compromise “Enabled Access” to 2,250 Tenants Where Microsoft Counted Sign-In Attempts and Cites “Over 1,600” Consoles Where Censys Counted 1,465, and a Microsoft Post It Cites Mistypes One Indicator in Its Hunting Query
The FBI’s November 13 FLASH on the ARTEX attack framework credits every indicator to an outside report, and checking five of its eleven notes finds all 24 addresses and three figures holding up, two...
The FBI’s FLASH-20261113-001, released on Friday, November 13, is a 13-page advisory marked TLP:CLEAR on ARTEX, the open-source, LLM-driven penetration-testing framework that, by the FBI’s account, a variety of malicious actors have used in intrusions since at least August. BleepingComputer reported it on November 15. Every indicator in the FLASH is credited to an outside report, so I read the primary pages behind five of its eleven notes (two Microsoft posts, CrowdStrike, Zscaler and Censys) and compared them with the FLASH’s text. All 24 addresses and three of the figures hold up. Two statements drift from their sources: a console count higher than the one in the cited post, and a tenant count described as access where Microsoft reported sign-in attempts. The Microsoft post behind the FLASH’s eight newest addresses also has a hunting query that mistypes one of them.
Table Of Content
This site has followed ARTEX since October: the South Korean bank attacks, the ConnectWise Cloud Backup compromise, Microsoft’s Storm-0940 report and the WorkOS breach. The FLASH is built from the same vendor reports that sit behind those earlier stories.
What the FLASH says
The advisory describes ARTEX as an autonomous, multi-agent platform written in Go that turns goals written in natural language into commands run by “super intelligence (SI) agents”. In its body the FLASH never writes “AI”: it says “agentic SI frameworks”, and the three places “AI” appears are inside the titles of the reports it cites. Its main claims:
- ARTEX has no model of its own. It connects to outside providers, which the FLASH lists as Anthropic Claude and OpenAI GPT among frontier models and DeepSeek, GLM, Grok and Kimi among open-source or Chinese-origin ones, and it says Chinese models are most prevalent on attacker-controlled instances (notes 7 and 9, Zscaler and Check Point).
- A developer called Autumn-27 released it on GitHub in July. The repository is gone, but forks, localized copies, archived copies and a commercial edition called ARTEX PRO keep it available.
- Two campaigns are named by description: attacks on South Korean financial institutions that exfiltrated personal information on about 67,000 people, and the ConnectWise Cloud Backup compromise. Some intrusions went from first targeting to privileged access in under three hours.
- Defenders are told to inventory forgotten internet-facing systems (partner portals, mobile backends, legacy services), restrict cloud administration endpoints, tighten application permissions and preserve identity logs.
What it leaves out is as informative. It names no bank, never mentions Claude Code, and says nothing of the Vapor Panda cluster or the South Korean police arrests that BleepingComputer’s article adds as background. It never uses the name Storm-0940; the only threat-actor name anywhere in it is Storm-3358, in an address table. Its table of observed techniques merges the observed behavior into one list with no actor and no source for any row, while the three indicator tables each carry a Source column.
Five of its eleven notes, checked
The FLASH’s eleven notes point to eight organizations: Akamai (two notes), Microsoft (two), Google (two), Censys, CrowdStrike, Zscaler, CERT-EU and Check Point. I read Microsoft’s October 21 and November 6 posts, CrowdStrike’s October 7 report, Zscaler’s report (dated October 14 in the FLASH’s note) and Censys’s November 5 post. I did not read the Akamai, Google, CERT-EU or Check Point notes, so claims that rest only on those, including the August start date, are unchecked here.
| FLASH statement | What the cited report says | Result |
|---|---|---|
| “some intrusions progressing from initial targeting to privileged access in under three hours” | Microsoft: “multiple instances where the time from initial targeting to privileged access was less than 3 hours” | Matches |
| “approximately 67,000 individuals” | Zscaler: “approximately 67,000 individuals” | Matches |
| 1,500 to 2,000 GitHub stars within 10 weeks | Censys: “reaching an estimated 1,500-2,000 stars within ten weeks of release” | Matches |
| 24 addresses, one domain and four hostnames | All 24 addresses appear in the report named for them (10 from CrowdStrike, 6 from Microsoft’s October 21 post and 8 from its November 6 post); the domain is CrowdStrike’s and the hostnames are Zscaler’s list | Matches |
| “over 1,600 publicly exposed ARTEX console instances” | Censys: “1,465 hosts running an exposed ARTEX console” as of November 4, about 650 of them “in two blocks of identically configured hosting”, “roughly 815 distinct-looking deployments”; no 1,600 anywhere in the post | Drifts |
| a compromise that “enabled access to more than 2,250 downstream tenants” | Microsoft: Storm-0940 “used ARTEX to attempt to sign-in to more than 2,250 tenants” | Drifts |
| repository made “unavailable on October 8, 2026” | Censys: the author posted a statement on October 8 and then said the project would go closed-source; the repository “has since been taken down”, with no date given | Partly |
The 2,250 tenants
The FLASH’s sentence is that the ConnectWise compromise “enabled access to more than 2,250 downstream tenants”. Microsoft’s sentence is that the actor “used ARTEX to attempt to sign-in to more than 2,250 tenants”, and its report adds that “the presence of an affected application or successful threat actor authentication does not necessarily indicate that customer data was accessed or exfiltrated”. Microsoft also writes: “While the majority of the targeting wasn’t successful, Storm-0940 achieved privileged cloud persistence at many organizations in a brief time.” The FLASH can be read as saying the compromise opened a path to those tenants, but “access” can easily be read as successful sign-ins. As our report on Microsoft’s post noted, the 2,250 is a count of targets, and Microsoft gave no count of compromises.
The 1,600 consoles
Note 4, Censys, is the only note attached to the FLASH’s console sentence. The live Censys post counts 1,465 hosts, and says “the ARTEX console does not require public-facing access, and overall global deployment is likely higher”. BleepingComputer repeats the figure as “more than 1,600”. If the FBI holds a newer count, the FLASH does not say where it comes from.
The newest indicators and a one-digit typo
The eight addresses the FLASH labels “Storm-3358 campaign” come from Microsoft’s November 6 post, which says that in October the actor began using ARTEX against the financial sector and calls it “a Cambodia-based intrusion cluster that primarily targets financial organizations in Southeast Asia”. Microsoft’s table gives each address a first-seen date and “Active as of 2026-11-06”. The FLASH keeps the second and drops the first.
The same Microsoft post also contains a hunting query with its own list of eight addresses, and the two lists disagree. The table has 178.16.54[.]129 (an ARTEX server first seen on October 19); the query has 178.15.54.129. The FLASH follows the table. The script below fetches the post and compares the two lists:
# ioc_diff.py (Python 3.10 or newer; needs: pip install requests)
# Compares the address table with the hunting-query address list inside one Microsoft post.
import html
import re
import requests
URL = ("https://www.microsoft.com/en-us/security/blog/2026/11/06/"
"artex-abuse-expands-as-cybercriminals-target-financial-sector/")
HEADERS = {
"User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 "
"(KHTML, like Gecko) Chrome/130.0.0.0 Safari/537.36",
"Accept-Language": "en-US,en;q=0.9",
}
page = requests.get(URL, headers=HEADERS, timeout=40).content.decode("utf-8", "replace")
text = re.sub(r"(?s)<(script|style).*?</\1>|<[^>]+>", " ", page)
text = re.sub(r"\s+", " ", html.unescape(text))
query = re.search(r"ioc_ip_addr = dynamic\(\[(.*?)\]\)", text).group(1)
in_query = set(re.findall(r"\d+\.\d+\.\d+\.\d+", query))
in_table = {ip.replace("[.]", ".") for ip in re.findall(r"\d+\.\d+\.\d+\[\.\]\d+", text)}
print("addresses in the table:", len(in_table))
print("addresses in the query:", len(in_query))
print("only in the table:", sorted(in_table - in_query))
print("only in the query:", sorted(in_query - in_table))
Run from a clean folder at 19:17 UTC on November 15, it printed:
addresses in the table: 8
addresses in the query: 8
only in the table: ['178.16.54.129']
only in the query: ['178.15.54.129']
Someone who pastes the query’s list into a SIEM searches for the wrong address and misses the server Microsoft lists. The registry records suggest the table is the correct one, which is my inference and not Microsoft’s statement: the range holding the table’s address belongs to the same Turkish provider as another listed address, while the query’s address sits in a small block registered to Vodafone Germany’s IP backbone. The records below come from the ARIN and RIPE RDAP services, read on November 15; they show who holds a range, not who operates the server.
| Address | Microsoft’s label | First seen | Holder of the covering range (RDAP) |
|---|---|---|---|
| 199.19.95[.]66, .67 and .68 | WISP C2 node | 2026-11-02 | 199.19.92.0/22, Netminders Server Hosting |
| 91.92.240[.]238 | ARTEX server | 2026-11-04 | 91.92.240.0/22, MGN Teknoloji Anonim Sirketi (Turkey) |
| 85.11.168[.]45 | ARTEX server | 2026-11-04 | 85.11.128.0/18, SOFCOMPANY Ltd (Bulgaria) |
| 85.11.167[.]52 | C2 server | 2026-11-04 | 85.11.167.0/24, TechTies Inc (Netherlands) |
| 178.16.54[.]129 | ARTEX server | 2026-10-19 | 178.16.52.0/22, MGN Teknoloji Anonim Sirketi (Turkey) |
| 31.77.229[.]35 | ARTEX server | 2026-10-28 | 31.77.229.0/24, network “guard”, registered 2026-10-24 (Netherlands) |
| 178.15.54.129 (query only) | not in Microsoft’s table | not applicable | 178.15.54.128/28, Vodafone Germany IP backbone block (HORGMB-NET) |
The page still carried both lists when I checked. Until Microsoft corrects it, build searches from the table.
What defenders can do this week
- Search from vetted, dated indicators. The FLASH itself tells readers to “investigate or vet IOCs prior to taking action” and to read addresses as “historically observed infrastructure within the activity window”. Microsoft says its three 187.15 addresses are VPN addresses and that “since they are VPN IPs, it is transient”, relevant to October 15 and 16.
- Handle the ConnectWise applications by application ID. The FLASH says to “search by application ID rather than display name” in every tenant where an affected ConnectWise application is or was authorized, and to keep the connections disabled “until ConnectWise provides formal restoration instructions”. ConnectWise’s Cloud Backup advisory, last updated on November 4 at 18:55 UTC, now describes how to re-authorize each Microsoft 365 tenant, with tenant administrators consenting to a new portal application, and still says: “Do not re-enable previously disabled applications unless ConnectWise specifically instructs you to do so.”
- Export identity logs now. The FLASH asks for a review of ConnectWise-related activity “starting from at least October 1, 2026” and to “preserve available Microsoft Entra sign-in and audit records”. Microsoft’s retention table gives Entra sign-in logs seven days on the free tier and 30 days on P1 and P2, and says of Graph activity logs that “data isn’t retained unless it’s archived to a storage account or integrated with analytics tools”. Thirty days before November 15 is October 16, and Microsoft dates the cross-tenant password spray to October 10 through 14, so default retention no longer reaches it. The Purview Unified Audit Log is separate, so check what your tenant keeps there.
- Treat the LLM hostnames as leads. The FLASH’s network indicator, calls to
api.anthropic.com,api.deepseek.com,open.bigmodel.cnorapi.x.ai, is Zscaler’s row for its South Korean campaign, whose configured backends were DeepSeek v4.1-flash, GLM-5.3 and Grok 4.6. The FLASH also names OpenAI GPT and Kimi but lists no hostname for them, and CrowdStrike says the Korean actor likely reached DeepSeek through a reseller domain (xcai[.]pro), so those calls would not hit the DeepSeek hostname at all. A hit from a server that has no reason to call those APIs deserves a look; a miss proves little. BleepingComputer says researcher MalwareMustDie released a detection pack on the same day; I have not tested it. - Report with the FBI’s checklist. The FLASH asks for eight kinds of information: ARTEX console artifacts, authentication and access logs, application and API activity, LLM infrastructure (gateways such as CLIProxyAPI and New API, resellers, billing anomalies), cloud and container activity, exploitation and persistence artifacts, attack characteristics (bursts, iterative failures, sub-agents overwriting each other) and loss and impact. Reports go to a local field office or IC3.
What the FLASH does not settle
- Where “over 1,600” comes from. The one note attached to it says 1,465.
- Whether the FBI has casework of its own. Every indicator row names a vendor, and the FLASH credits no FBI source.
- How many tenants were compromised. The FLASH gives no number, and Microsoft’s October 21 text says only “many organizations”.
- Whether the six notes I did not read (Akamai twice, Google twice, CERT-EU, Check Point) support the claims attached to them.
No ARTEX code, repository or exploit material was opened or linked for this report, and nothing was sent to any listed address. The registry lookups went only to public RDAP services.







No Comment! Be the first one.