ARTEX: Threat Landscape to Detections Today
Written by Tim Peck · Director of Threat Research

Sources: public community rules and public Threat Reports on detections.ai, and the public research linked below. Referenced figures run from 7 Oct to 16 Nov 2026. All figures are aggregates.

The back story and how it rose in popularity
In one intrusion that Google's incident responders investigated, an ARTEX agent redirected MySQL's general log to write a JSP web shell, hit a compile error, and rewrote the file under new names until one ran: cl1217.jsp, f1219a.jsp, z1219.jsp. In another case, Microsoft counted more than 151,000 failed logons from a Storm-0940 password spray that never produced a sign-in.
Both details come from the same habit. ARTEX tries many things quickly and reads the errors, so an operation leaves a dense trail of failed attempts in the victim's logs.
ARTEX ships without a model. The operator supplies an API key for Claude, GPT, DeepSeek, GLM, Grok or Kimi, then a target and an objective. A planner agent splits the objective into tasks, worker agents carry them out with ordinary tools such as nmap and curl, and findings go into a shared graph that the planner reads before it picks the next step.
It is written in Go, with a web console and PostgreSQL behind it. The FBI says DeepSeek's Flash models are the most common choice on attacker-run instances.
A developer using the handle Autumn-27 published ARTEX on GitHub on 26 July 2026. On 8 October, a day after CrowdStrike tied it to breaches at South Korean financial firms, the developer said the project would go closed source, and the repository went offline. Copies were already circulating. Censys found 1,465 hosts exposing an ARTEX console in early November, about 815 of them distinct deployments. Cloudflare says ARTEX and its forks made up more than 60% of the AI pentest orchestrators it knew of on 12 November.
Three named clusters are on record for using this tool. Microsoft ties Storm-0940, a China-aligned actor, to a campaign against hundreds of organizations from September, in which stolen ConnectWise Cloud Backup application credentials were used to attempt sign-in at more than 2,250 downstream tenants. CrowdStrike reports VAPOR PANDA targeting more than 200 organizations in at least 14 countries over 11 days from 17 October, with post-exploitation confirmed at nine. Microsoft's Storm-3358 is financially motivated and aims mostly at financial firms in Southeast Asia.
The FBI's FLASH of 13 November says some intrusions reached privileged access in under three hours. The reports cited here describe no zero-day. The agents got in through SQL injection, missing authentication, unsigned SAML assertions, admin passwords hardcoded in JavaScript bundles and keys that were never rotated.

Figure 2: note: industry dates are publication dates unless the entry says otherwise. detections.ai times are UTC and refer to public Threat Reports and public community rules.
Where an ARTEX operation shows up
The harness leaves evidence in two places. One is the operator's host, where the console, the agents and their tools run. You only see that side if someone runs ARTEX inside your network or from a server of yours. The other is your own telemetry, where the workers' requests and commands land.
How the harness works, and how many public rules look at each part

Figure 3: diagram showing 50 public rules on detections.ai as of 16 Nov 2026, counted once each and placed by what the query looks at.
Twenty-seven of the 50 rules cover two stages: identity and cloud abuse, and what happens on a host after a foothold. That is also where the vendor reports put the damage.
Detections you can reliably use today
We hand picked 9 notable detections and these are the ones we would deploy first, in kill-chain order. Each entry says what the rule depends on. Rule pages need a free login.
- Storm-0940 Password Spray Hunt: Many Accounts From One Address, Scripted Clients and Campaign AddressesKQL · Initial access · Keys on behavior, plus indicators · Needs EntraIdSignInEventsEntra sign-ins per address in 10-minute windows: at least 10 accounts failing, at least 30 failures and a failure ratio of 0.5 or more. It also returns any failed sign-in from four campaign addresses, and flags Python and curl clients. The rule states its own gap: a spray spread across many addresses.
- Storm-0940 ARTEX password spray followed by OAuth app credential abuseKQL · Identity and cloud · Keys on behavior, plus indicators · Needs SigninLogs and AuditLogsA spray (20 accounts and 50 failures from one address in 10 minutes, or any failure from three known Storm-0940 addresses), a successful sign-in from that address, and an application credential, consent or role change from it within 6 hours of the spray starting. Bulk credential updates from a shared corporate address can match.
- Service Principal Self-Assignment to Directory RoleKQL · Identity and cloud · Keys on behavior · Needs AuditLogsAn application adding itself to a directory role, or granting itself an app role, that it has not assigned itself in the previous 7 days. Huntress saw the directory-role step in about 50 tenants on 15 October.
- Storm-0940 Key Vault Hunt: One Identity Reading Secrets Across Several Vaults Beyond Its BaselineKQL · Identity and cloud · Keys on behavior · Needs CloudKeyVaultEventsOne identity reading or listing secrets in 3 or more Key Vaults, with 10 or more reads in a day, across more vaults than on any day in the previous two weeks.
- ARTEX and AntiFox AWS Agentic Pentest HuntKQL · Identity and cloud · Keys on behavior, plus indicators · Needs AWSCloudTrail in SentinelAt least 50 read calls across 8 or more services from one address in 30 minutes with at least 10 denied, CloudWatch log searches for AKIA or ?password, and placeholder S3 object probes. It also matches published addresses and user agents naming ARTEX or AntiFox, a fork of ARTEX.
- ARTEX - Database engine writing a server-side script file (web shell planting) KQL · Execution and post-exploitation · Keys on behavior · Needs DeviceFileEvents mysqld, mariadbd or postgres creating, modifying or renaming a server-side script file such as .jsp, .php or .aspx, with web paths flagged. This is the MySQL log technique from Google's report.
- ARTEX - Web | DB server lineage spawning curl to pull scripts or beacon to webhook.siteKQL · Execution and post-exploitation · Keys on behavior · Needs DeviceProcessEventscurl or wget under a shell under a web or database server process, fetching a script or calling webhook.site.
- Web server process spawning shell after web-served file write (web shell chain)KQL · Execution and post-exploitation · Keys on behavior · Needs DeviceFileEvents and DeviceProcessEventsLinux: PHP, Apache or nginx writes a script into a web root, and a web process on that host spawns a shell running curl, wget, nc, chmod or base64 in the same clock hour.
In addition to these, detections.ai hosts more than 50 rules that are in KQL (47), Suricata (2) and XQL (1).
When the indicators change…
The most-copied ARTEX rule on detections.ai is an indicator sweep. It checks inbound connections against a third-party feed of IP addresses tied to ARTEX and CyberStrikeAI, and most engineers who copied any of these rules copied that one. It answers the first question every team asks, which is whether known ARTEX infrastructure has touched them.
That answer ages quickly. Censys watched 10 open directories on ARTEX hosts close and 15 new ones appear on other hosts between 3 and 5 November. MalwareMustDie's detection pack, the most complete public rule set for ARTEX we have read, stops alerting on its own IP indicators after 90 days and labels them context for retro-hunting.
The behavior changes much less. Five patterns recur across the vendor reports, and each one shows up in a log source most teams already collect.
- Cloud permission probing in burstsGoogle saw AWS DryRun calls with placeholder IDs (27 events in 23 seconds), an 8-second ARM deployments/validate sweep across 11 resource types, and a blob written and deleted within 400 ms.
- Fail, correct, retry on identityMicrosoft saw ARM calls return 400 and then 200 with only the api-version changed, 28 to 80 seconds apart, and AADSTS7000215 followed by a success for the same application and address. Huntress saw retries 3.5 and 7.1 seconds after a failure, and service principals assigning themselves directory roles.
- Servers behaving like clientsGoogle saw curl on a Windows database server pull generated scripts and post results to 491 rotating webhook.site URLs, and a database engine write .jsp files. Microsoft saw netcat and curl fetch open-source RATs.
- Reading the application before attacking itWiz saw requests for source maps, /env.js and OpenAPI documents, then calls to routes named only in those files. It also saw more than 60 SSO route variants in two minutes, and logins one second apart for 20 to 60 minutes with a new user agent per burst.
- Leftovers from self-checkingGoogle found probe files named artex-probe- plus a five-digit number, accounts containing ArtexProbe, and test emails sent to more than 1,300 internal users through a notification feature.
Speed on its own is a weak signal, because inventory scripts and CI jobs are fast too. MalwareMustDie ships each timing pattern as informational and raises it only in combination with something else. Pair a burst with an action ordinary automation doesn't take: denied calls across many services, a credential added to an application, a shell under a web worker.
Harness-specific strings are worth adding as a second signal. Reported ones include the user agents norma/0.4 and artex-enrich/1.0, an HTML title beginning ARTEX on port 8787, command lines containing artx and cloud roles prefixed artx. Each is a default or one operator's habit. CrowdStrike's artx markers come from the VAPOR PANDA campaign, and a fork can change a user agent in one line of configuration.
The harness can also run inside your network, as an unsanctioned test or on a compromised server used as a launch point. In that case look for a console on port 8787 with an unauthenticated HTTP proxy on 8788 beside it, a container image with artex in its name, scanners starting inside a container, and servers calling LLM APIs they have no reason to call. Two Docker images are known to be backdoored: tomicdd/artex and congdonglinux/artex.
Demand arrived four weeks after the first rules
The first search for artex on detections.ai came on 7 October. The first rule was published just a few hours later, in a batch of six rules and a Threat Report shared to five community groups.
Interest stayed low for four weeks. Then on the 5 and 6 November popularity rose again with hundreds of searches driving to both detections and intel. Censys, Check Point, CERT-EU and Microsoft all published in those two days, and the indicator sweep went up on the 5th.
Searches, copies and published rules, 7 October to 16 November

Figure 4: distinct people per day - Bottom: distinct public rules published so far, with a tick for each day a public Threat Report was published. Vertical lines mark industry reports. 16 November is a partial day. Use the arrow keys on the chart to step through days.
Working this on detections.ai
Start from a Threat Report. Over a dozen public reports mention ARTEX and are available now on detections.ai. Here are a few that scored at least a 9/10 on our intel scoring metric:
- ARTEX and Offensive Agentic AI Intrusions, from Wiz and Google
- VAPOR PANDA AI-Driven ARTEX Global Espionage Campaign, from CrowdStrike
- Storm-0940 AI-Driven ConnectWise Supply Chain Compromise, from Microsoft
- Adoption of Agentic AI Offensive Frameworks in Intrusions, a cross-vendor roundup
- ARTEX AI-Driven Offensive Framework Exploitation Analysis, published today
Then search by actor as well as by tool. artex returns 23 of the 50 rules, and Storm-0940 returns about a dozen. For Sigma, YARA, Snort and osquery content written for ARTEX, the MalwareMustDie pack is published under a license that allows reuse with attribution.
AI Assistant will draft a detection from any of these reports for the stack you run. The Library's ATT&CK coverage map then shows which of the techniques in the report your saved detections already cover.
For teams on the enterprise tier, Nightwatch reads each new report as it arrives and answers two questions for your environment: are we impacted, and are we covered. It hunts your connected telemetry for the first and checks your own detection library for the second. Nothing reaches your SIEM without you.
Limits and sources
- The rules are new. 44 of the 50 are less than four weeks old and five were published today, so test before alerting. 31 came from detections.ai's AI pipeline and 19 from community members.
- Read the logic before you deploy. The stage and dependency labels are our reading of each query, and a few titles promise a condition the query doesn't test.
- Vendor figures disagree, and we have kept each with its source. Null Security traced the seven claims in the FBI FLASH's summary to the reports it cites and found that five match. The FLASH describes Microsoft's 2,250 sign-in attempts as access, and credits "over 1,600" exposed consoles to Censys, which counted 1,465. The figure of about 67,000 people affected in South Korea appears in Zscaler's report alone.
- Huntress, which worked the ConnectWise intrusions directly, rates LLM involvement as low to medium on its own data. The ARTEX attribution in that case is Microsoft's.
- We have not run ARTEX or reproduced any intrusion. Harness details come from the cited reports and the documentation of a public fork.
- 16 November is a partial day. Usage counts cover logged-in members and exclude detections.ai staff.
Appendix
Rules: Metrics for this report were gathered from published rules in community groups and counted once per distinct title and logic. A rule names ARTEX when the word is in its title, description or logic. A rule is counted as written from the same reports when it cites an ARTEX source article or carries the artex tag without using the word. One rule about a defensive decoy tool is left out. Usage: distinct people in product analytics from 8 October, matched to those rules by id. Search: distinct people whose query was artex, artex ai or artex pro.
Research: FBI FLASH-20261113-001, Microsoft on Storm-0940, exposed secrets and Storm-3358, CrowdStrike on South Korea and VAPOR PANDA, Google Threat Intelligence, Zscaler ThreatLabz, Wiz, Censys, Check Point Research, Huntress (16 Oct, 26 Oct), Cloudflare, Cisco Talos, CERT-EU, GitGuardian, Null Security, MalwareMustDie.


