Most Publishers on ARTEX Had Their Own Evidence, but Some Numbers Drifted

We read every report in Asterism's panel that named ARTEX, the open-source penetration testing framework run by large language models, from 19 Sep to 15 Nov 2026. We coded the evidence each one brought, and traced the claims in the FBI's FLASH of 13 Nov back to the reports it cites.

Marking
TLP:CLEAR
Type
Research
Author
Null Sec team
Published
· 7 min read
Applies to
Reports from 19 Sep to 15 Nov 2026
Data
Counts of public reporting in Asterism's panel, as of 16 Nov 2026
An engraving of rings of ripples spreading across a calm lake from a single fallen stone.

Of the 38 organizations in Asterism's panel that reported on ARTEX from 19 Sep to 15 Nov 2026, 23 published evidence of their own, in 33 of the 68 reports. Some of the numbers repeated about it drifted from their sources. The FBI's FLASH of 13 Nov footnotes every claim in its summary to one of 11 public reports,1 and five of its seven summary claims match the reports they cite. Two don't: it counts Microsoft's sign-in attempts against more than 2,250 tenants as access to them, and it credits "over 1,600" exposed consoles to a Censys report that counted 1,465.

Key Findings

  • Of the 68 ARTEX reports, 33 (49%, 95% CI 37% to 60%) brought the publisher's own evidence: telemetry, incident response, scans for exposed consoles, or a reading of the code. By organization, 23 of 38 did, about 3 reports for each one.
  • Of the seven claims in the summary of the FBI's FLASH, five match a report it cites. Microsoft reported sign-in attempts against more than 2,250 tenants, and the FLASH says the compromise enabled access to them. Its "over 1,600" exposed consoles is credited to Censys, which counted 1,465.
  • The vendor reports that repeated Microsoft's tenant figure kept its meaning: 12 of the 16 repeats describe sign-in attempts. The FLASH is the only one that describes access.
  • The counts of exposed ARTEX consoles in circulation run from 600 to 2,500, and they count different things: IP addresses, hosts, live panels, the US alone, command servers.
  • The figure of about 67,000 people whose data was taken in South Korea is in one of the three reports the FLASH cites for it, with no source given. The earliest count we found, about 66,000, came from Korean press reports that named none either.

Most Organizations Brought Their Own Evidence

ARTEX reports by evidence · 19 Sep to 15 Nov 2026
  • whisker, the interval
Share of ARTEX reports0%10%20%30%40%50%first-hand, ARTEX in usefirst-hand, ARTEX in use: 29% (20% to 41%), n = 2029%n = 20first-hand, ARTEX linkcreditedfirst-hand, ARTEX link credited: 6% (2% to 14%), n = 46%n = 4first-hand, tool or exposurefirst-hand, tool or exposure: 13% (7% to 23%), n = 913%n = 9relay of othersrelay of others: 25% (16% to 36%), n = 1725%n = 17passing mentionpassing mention: 27% (17% to 38%), n = 1827%n = 18Share of ARTEX reports0%20%40%60%first-hand, ARTEX in usefirst-hand, ARTEX in use: 29% (20% to 41%), n = 2029%n = 20first-hand, ARTEX link creditedfirst-hand, ARTEX link credited: 6% (2% to 14%), n = 46%n = 4first-hand, tool or exposurefirst-hand, tool or exposure: 13% (7% to 23%), n = 913%n = 9relay of othersrelay of others: 25% (16% to 36%), n = 1725%n = 17passing mentionpassing mention: 27% (17% to 38%), n = 1827%n = 18

as of 16 Nov 2026Measures reporting, not activity.Download CSV

Data table classes.csv
The data behind Fig. 1, from classes.csv.
classsharelohin
first-hand, ARTEX in use0.2940.1990.41120
first-hand, ARTEX link credited0.0590.0230.1424
first-hand, tool or exposure0.1320.0710.2339
relay of others0.2500.1620.36417
passing mention0.2650.1740.38018
Fig. 1Share of ARTEX reports by the evidence they carried, with 95% confidence intervals. A report takes the highest class that applies; the first three are first-hand.Plate I

A first-hand report carries something the publisher collected itself: an incident response case, its own telemetry, a scan for exposed consoles, a reading of the code. Of the 68 reports, 33 did (49%, 95% CI 37% to 60%). Counting only reports whose own evidence ties the activity to ARTEX, not those crediting the link to someone else, gives 43%. The rest relayed others' findings (17 reports) or named ARTEX in passing (18).

By organization, 23 of the 38 published at least one first-hand report (61%, 95% CI 45% to 74%), about 3 ARTEX reports for each. That is a dense primary record for one tool in six weeks. It is also the record of our panel, which is security vendors and researchers with one CERT and one government agency, and no news outlets. The news coverage around ARTEX is outside it.

ARTEX reports a week · 5 Oct to 15 Nov 2026
  • first-hand
  • relay
  • mention
Reports a weekWeek starting05101520255 Oct, first-hand: 15 Oct, mention: 112 Oct, first-hand: 312 Oct, mention: 219 Oct, first-hand: 519 Oct, relay: 319 Oct, mention: 326 Oct, first-hand: 926 Oct, relay: 126 Oct, mention: 32 Nov, first-hand: 82 Nov, relay: 52 Nov, mention: 79 Nov, first-hand: 79 Nov, relay: 89 Nov, mention: 29 Nov5 Oct12 Oct19 Oct26 Oct2 Nov2017Reports a weekWeek starting01020305 Oct, first-hand: 15 Oct, mention: 112 Oct, first-hand: 312 Oct, mention: 219 Oct, first-hand: 519 Oct, relay: 319 Oct, mention: 326 Oct, first-hand: 926 Oct, relay: 126 Oct, mention: 32 Nov, first-hand: 82 Nov, relay: 52 Nov, mention: 79 Nov, first-hand: 79 Nov, relay: 89 Nov, mention: 29 Nov5 Oct19 Oct20

as of 16 Nov 2026Measures reporting, not activity.Download CSV

Data table weekly.csv
The data behind Fig. 2, from weekly.csv.
weekkindreports
2026-10-05first-hand1
2026-10-05relay0
2026-10-05mention1
2026-10-12first-hand3
2026-10-12relay0
2026-10-12mention2
2026-10-19first-hand5
2026-10-19relay3
2026-10-19mention3
2026-10-26first-hand9
2026-10-26relay1
2026-10-26mention3
2026-11-02first-hand8
2026-11-02relay5
2026-11-02mention7
2026-11-09first-hand7
2026-11-09relay8
2026-11-09mention2
Fig. 2ARTEX reports a week, by the evidence they carried. Weeks start on Monday; the last week includes the FLASH.Plate II

First-hand reports came early and kept coming. Relay grew with the attention: none in the first two weeks, then 8 of the 17 reports in the week of the FBI's FLASH, the FLASH among them.

The FLASH's Seven Claims, Checked

The FLASH is a synthesis. It presents nothing as the FBI's own observation: the claims in its summary carry numbered notes, and each indicator row names the vendor that published it. Those notes are what made it the easiest report in our set to check. Of its 11 notes, 7 are reports in our panel. We read CERT-EU's on its site, and couldn't find the two Akamai posts or Mandiant's companion post. Every claim below cites at least one report we read.

Table 1The seven claims in the summary of FLASH-20261113-001, the notes it gives for each, and whether those reports say the same, read on 16 Nov 2026. What each cited report says is in flash-claims.csv.
resultthe FLASH's claimnotes
matchesused since at least August 2026[1] to [5]
matchesprivileged access in under three hours[2]
in one of threeabout 67,000 people in South Korea[6] [7] [8]
stronger than the sourceaccess to more than 2,250 tenants[2]
matchesrepository removed in October[3]
in one of twoforks and a commercial ARTEX PRO[3] [10]
not in the cited reportover 1,600 exposed consoles[4]

Attempts became access. Microsoft's report of 21 Oct says that after ConnectWise disclosed unauthorized access to its customers' tenants, Storm-0940 (Microsoft's name for the actor) used ARTEX to reach more than 2,250 tenants at up to 2,900 sign-in attempts a minute.2 The same report describes several intrusions that reached privileged access in under three hours, a separate claim the FLASH carries correctly. The FLASH joins the two into one sentence: a supply chain compromise "that enabled access to more than 2,250 downstream tenants". The number is Microsoft's, and in Microsoft's report it counts attempts.

A count from a different report. For "over 1,600 publicly exposed ARTEX console instances", the FLASH's note [4] is Censys's "ARTEX: Agentic Asbestos".3 Censys counted 1,465 hosts with an exposed console as of 3 Nov (4 Nov in its summary), and about 815 once two blocks of identically configured hosting are set aside. Its page showed no later figure when we read it on 16 Nov. Check Point's report of the same day, which the FLASH cites as note [9] for other claims, gives 1,608 live panels.4 That is the nearest match, though the FLASH doesn't say it used it.

Two more claims rest on one of their notes. Microsoft's post of 6 Nov, cited with GTIG's for the forks and ARTEX PRO, mentions neither.5 Of the three notes for the South Korean victim count, only Zscaler's gives it (see below).

Microsoft's Tenant Figure in Other Reports

Repeats of the tenant figure · 21 Oct to 15 Nov 2026
Reports024681012sign-in attempts against themsign-in attempts against them: 1212reached themreached them: 11Microsoft's count, no verbMicrosoft's count, no verb: 11swept OAuth tokens across2,200swept OAuth tokens across 2,200: 11enabled access to themenabled access to them: 11Reports051015sign-in attempts against themsign-in attempts against them: 1212reached themreached them: 11Microsoft's count, no verbMicrosoft's count, no verb: 11swept OAuth tokens across 2,200swept OAuth tokens across 2,200: 11enabled access to themenabled access to them: 11

as of 16 Nov 2026Measures reporting, not activity.small sample (n = 16)Download CSV

Data table wording.csv
The data behind Fig. 3, from wording.csv.
wordingreports
sign-in attempts against them12
reached them1
Microsoft's count, no verb1
swept OAuth tokens across 2,2001
enabled access to them1
Fig. 3Reports that repeated Microsoft's ConnectWise tenant figure, by how each one described what happened at those tenants. The FLASH is the last bar.Plate III

Most repeats stayed close to Microsoft's sentence: 12 of 16 describe sign-in attempts against the tenants, and all but one credit Microsoft. One weekly round-up wrote that Storm-0940 swept OAuth tokens across more than 2,200 tenants.6 Only the FLASH turned the attempts into access. Each report's wording is in wording-2250.csv.

Six Counts of Exposed Consoles

Table 2The counts of exposed ARTEX consoles in the reports in our panel, and what each one counted, as each report states it. The details are in consoles.csv.
countpublisherwhat it counted
600AhnLab, 13 OctIP addresses hosting ARTEX
1,465Censys, 5 Novhosts with an exposed console
1,608Check Point, 5 Novlive panels
916SOCRadar, 12 Novlive consoles in the US alone
2,500Cisco Talos, 13 Novcommand server addresses
over 1,600FBI, 13 Novexposed consoles, credited to Censys

None of these is wrong by itself. They measure different things on different days, and a number quoted without its definition is how they collide. Check Point's 1,608 live panels come from 2,557 that host or once hosted one, and elsewhere the same report says more than 1,800. SOCRadar's 916 covers the US alone, from 1,295 candidates,7 and Talos's 2,500 counts command server addresses, not consoles.8 AhnLab, the first to publish one, scanned on 13 Oct, before most of the attention.9

A Victim Count Without a Source

The earliest report in our panel with the figure of about 67,000 individuals is Zscaler's of 14 Oct, which gives no source for it.10 CrowdStrike's report of 7 Oct, which the FLASH cites first for the figure, has no count, and says the number of organizations affected was still unconfirmed.11 CERT-EU's post has none either.12 Hunt[.]io links a Korea Herald report of 7 Oct, which puts the exposure at about 66,000 individuals and 2,200 corporate records, credited only to "reports".1314 The figure may be right. The public record we could read doesn't show where it comes from.

Limits of This Report

  • This measures reporting, not activity. First-hand means the publisher says the evidence is its own. We didn't verify anyone's telemetry, and a first-hand report can be wrong.
  • Our panel is mostly English-language security vendors and researchers, with one CERT, one government agency and no news outlets. Publishers that opted out of AI use are invisible to us.
  • The window ends on 15 Nov, two days after the FLASH. Reports citing it were still arriving on the as-of date, and they aren't counted.
  • ARTEX was released in July. Reporting before Asterism's coverage began on 18 Sep is missing; the first ARTEX report in our panel is CrowdStrike's of 7 Oct.
  • Coding a report's evidence is a judgment. A second coder agreed on 22 of 25, and every disagreement was at the edge of first-hand: one sentence of a publisher's own telemetry, or a report restating its own earlier research.
  • We checked the FLASH against the cited reports as they read on 16 Nov. Publishers edit posts: Microsoft's report of 21 Oct now says "attempt to sign-in to" where Asterism's copy says "reach". Neither says access.
  • We couldn't find three of the FLASH's notes, the two Akamai posts and Mandiant's companion post. No claim we checked rests on them alone.

Method

The definitions of a report and an organization are those of How Null Sec Counts. An ARTEX report is a report dated 19 Sep to 15 Nov 2026 (UTC) whose text names ARTEX and which is about the framework, its forks or ARTEX PRO. A full-text search found 69; we dropped one where "artex" was a folder name. Each source here is its own organization (organizations.csv).

We read each report in full and coded the evidence it carried, from a rubric written before the reading. The classes, highest first, with a report taking the highest that applies:

  • first-hand, ARTEX in use: the publisher's own observation of ARTEX used against real targets, from its incident response, telemetry, sensors, or an investigation;
  • first-hand, ARTEX link credited: the publisher's own telemetry of an intrusion, with the link to ARTEX credited to someone else;
  • first-hand, tool or exposure: the publisher's own work on ARTEX itself, such as scans for exposed consoles or a reading of the code;
  • relay of others: ARTEX covered at length, with every factual claim from other publishers;
  • passing mention: ARTEX named in passing.

A report restating its own organization's earlier first-hand evidence counts as first-hand. The second class and that rule were settled during the coding, before the counts. A second coder, given only the rubric, coded a random sample of 25 reports and agreed on 22 (Cohen's kappa 0.85); we decided the disagreements by the rubric, and validation.csv lists them.

The share of first-hand reports holds up to three changes: without the 14 round-ups it is 59%; with only reports where ARTEX is the subject or covered at length, 61%; and with the window ending on 12 Nov, before the FLASH, 48%. Shares have Wilson intervals.

The claims are the seven in the FLASH's summary, chosen before we counted anything. For each report we recorded whether it states each claim and whether it names a source (claims.csv), and full-text queries cross-check the reading (queries.aql). We read the FLASH's notes in the TLP:CLEAR PDF, the cited reports in Asterism, and on 16 Nov, the cited reports' pages on their publishers' sites.

About the Data

As of
Window
19 Sep to 15 Nov 2026
Coverage since
Panel
182 sources reported in the first two weeks of the window, 179 in the last two; the data counted here come from 38 organizations
Marking
TLP:CLEAR
Left out
Undated reports, marketing and product news, publishers who opted out of AI use, anything above TLP:CLEAR, and reports where "ARTEX" means something else (one). Round-ups are counted and marked.

This measures reporting, not activity. It counts what was published.

Null Sec sells Asterism, the platform these data come from.

Method: How Null Sec Counts

Data and Queries

The files behind every figure and number in this report, as published. Each SHA-256 lets you check you have the same bytes.

  • classes.csvThe data behind Fig. 1.247 bytes · SHA-256 2251c6acb093dc6d170a502ec75968f93c0cf5afd27ca8d891de1ee6ab5ad8d3
  • weekly.csvThe data behind Fig. 2.402 bytes · SHA-256 0252b22361d97a2da7d1b88fc1caaf45bea80a58ec33189fbbe19f007788a956
  • wording.csvThe data behind Fig. 3.156 bytes · SHA-256 69b660477685ecd869e3036246a060fb93effb846b768520adcbfacfd8d857cc
  • claims.csvSupporting data.558 bytes · SHA-256 4b989c012c8f90243ce20db60709577258a6c3232bd05da21893dbefeb86f150
  • consoles.csvSupporting data.611 bytes · SHA-256 6f16a367c2d528e8e83462e50aec4ac3f84891d4393a62320fdef81b0bed9a61
  • flash-claims.csvSupporting data.1,009 bytes · SHA-256 2c99c3983bbe8339f56e1c29dd2aa98725a5c0e998396cd3802c051b5cd59c33
  • organizations.csvSupporting data.936 bytes · SHA-256 f4b8ac1f40f9db3a39b174676a10d0581ba530b83c34bebada4269b6e98e8fe8
  • reports.csvSupporting data.17.4 KB · SHA-256 7bf2da73784934a97783ff873392b4d0674dc9b03adac50e0c97b03286cf6f6b
  • validation.csvSupporting data.4.5 KB · SHA-256 4bbe5c4ea1c0bf3ac0475f45b4c34aec40a08578ac800cc215cf0e77b9bd843d
  • wording-2250.csvSupporting data.1.4 KB · SHA-256 9e85c9fe6bcd628b3473d73f4f6076fe2c5428525eff9967edd1af87181acee3
  • queries.aqlThe exact queries, with their windows and the dates they ran.2.4 KB · SHA-256 c915cdb6a9e6cbf173b79c2e81f1685213f1b910ebba98ea750eab984e44e8ad
  • sources.csvEvery report cited or counted, with its publisher, date, URL and TLP.15.9 KB · SHA-256 4a1bf8c210b46c02a0b3e5d3cce3b5b7d12139e7c1908bdaa6664f3544b592b6

References

  1. FBI, "Threat Overview of ARTEX Autonomous Penetration Testing Framework", FLASH-20261113-001, 13 Nov 2026. TLP:CLEAR. www.ic3.gov/CSA/2026/261113.pdf ↑
  2. Microsoft Threat Intelligence, "AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise", 21 Oct 2026. www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/ ↑
  3. Censys, "ARTEX: Agentic Asbestos", 5 Nov 2026. censys.com/blog/artex-agentic-asbestos/ ↑
  4. Check Point Research, "Hacking on Autopilot: Inside ARTEX Cybercrime Campaigns", 5 Nov 2026. research.checkpoint.com/2026/hacking-on-autopilot-inside-artex-cybercrime-campaigns/ ↑
  5. Microsoft Threat Intelligence, "ARTEX abuse expands as cybercriminals target financial sector", 6 Nov 2026 (dated 7 Nov in Asterism). www.microsoft.com/en-us/security/blog/2026/11/06/artex-abuse-expands-as-cybercriminals-target-financial-sector/ ↑
  6. RedSheep Security, "Weekly Threat Intel Report, 2026-W43: 19 to 25 October 2026", 25 Oct 2026. www.redsheepsec.com/trends/weekly-report-2026-w43 ↑
  7. SOCRadar, "ARTEX: 916 Exposed AI Pentest Consoles Mapped in the U.S.", 12 Nov 2026. socradar.io/blog/artex-ai-pentest-consoles-us/ ↑
  8. Cisco Talos, "ARTEX and beyond: AI-enabled penetration testing platforms", 13 Nov 2026. blog.talosintelligence.com/artex-and-beyond/ ↑
  9. AhnLab ASEC, "The Threat of AI-Driven Autonomous Penetration Testing, as Examined by ARTEX", 13 Oct 2026. asec.ahnlab.com/en/95845/ ↑
  10. Zscaler ThreatLabz, "Threat Actor Abuses Agentic AI Pentesting Framework to Target South Korean Financial Institutions", 14 Oct 2026. www.zscaler.com/blogs/security-research/threat-actor-abuses-agentic-ai-pentesting-framework-target-south-korean ↑
  11. CrowdStrike, "Unknown Threat Actor Uses AI-Driven ARTEX to Target South Korean Finance", 7 Oct 2026. www.crowdstrike.com/en-us/blog/unknown-threat-actor-uses-artex-to-target-south-korean-finance/ ↑
  12. CERT-EU, "From Penetration Testing to Cyber Threat: Beware of ARTEX", 5 Nov 2026. cert.europa.eu/blog/from-penetration-testing-to-cyber-threat-beware-of-artex ↑
  13. Hunt[.]io, "ARTEX Exposed: Infrastructure Behind the South Korea Bank Hack", 4 Nov 2026. hunt.io/blog/artex-south-korea-bank-infrastructure ↑
  14. Park Yuna, The Korea Herald, "South Korean police launch major probe as suspected AI hacks sweep through banks", 7 Oct 2026, republished by Asia News Network. asianews.network/south-korean-police-launch-major-probe-as-suspected-ai-hacks-sweep-through-banks/ ↑
TLP:CLEAR Share freely, subject to standard copyright rules.Corrections: none so far

Spot an error? Write to us and we'll fix it here, with a dated note.