Summary
A trojanized Docker image of ARTEX, an open-source AI-driven penetration testing console, was found handing out unauthenticated root shells: a single HTTP request to the console's own login endpoint runs any command and returns the output, no credentials required. Once the backdoor was public, I ran an independent census against its fingerprint using Censys to measure how exposure was actually changing.
Over four days, the count of live backdoored consoles fell by roughly 59%. But the population didn't just shrink: one cluster of four hosts came online two days after the disclosure, on non-default ports: evidence of an operator who kept deploying a tool they knew was compromised, not someone unaware of it.
Update, 13 November 2026
Three things this report did not cover when it was published on 9 November.
The India link. On 4 November, the same day as BadByte's disclosure, CERT-In published advisory CIAD-2026-0058 on a threat actor it calls Crafty Falcon, which is attacking Indian banking and financial firms. The advisory names ARTEX among the tools that actor uses, alongside PentAGI, Cairn and CyberStrikeAI, as reported by The Hacker News. This is context, not attribution: no public reporting ties the backdoored image to Crafty Falcon. The two share a tool, and that is all the evidence shows.
The baseline. Press coverage quotes 62 backdoored consoles. That is BadByte's count of every console seen between 2 and 4 November. Of those, 46 were still serving the backdoored build at 09:55 UTC on 4 November. The 59% figure in this report compares that 46 with my 19, a point-in-time count against a point-in-time count.
The second image and the repositories. congdonglinux/artex, listed in the indicators, is a byte-identical copy of the backdoored image, and Docker removed both on 5 November. BadByte also counts 637 GitHub repositories whose compose file still points at the image, most of them forks of a single repository. Since the removal, docker compose up in those repositories fails instead of pulling the backdoor, so they are stale pointers, not a working install path. The exposure that remains is on hosts that pulled the image before 5 November, which is what this census measures. I have not recounted the repositories myself.
Background
ARTEX pairs a web console with LLM agents that scan, exploit, and write up findings against a target with minimal human direction. It had already been used maliciously in the wild. CrowdStrike linked ARTEX-driven activity to intrusions at South Korean banks in late September and early October, and to a China-nexus campaign against more than 200 organizations in 14 countries. On 8 October 2026, the original author posted that the project had been abused, announced it was going closed-source, and took the repository down.
Within hours, a backdoored fork filled the vacuum: a new Docker Hub account pushed a copy of the image with one function added, and a new GitHub account pushed a copy of the source pointed at that image, presented as the "maintained successor." The source itself was clean, so a code review would find nothing. The backdoor lived only in the prebuilt binary.
The binary diffing, disassembly, and initial disclosure are credited to BadByte's original investigation, published 4 November 2026; see their report for the full technical breakdown, YARA rules, and detection guidance. The census and timing analysis below are original to this report.
Methodology
Every backdoored instance serves a fixed UI build ID in its HTML, 3aC1db3gGU08O4UHbAt4J, regardless of hostname or port. This is the same fingerprint BadByte used, built from the same underlying scan data (both this census and BadByte's credit Censys as the dataset source), so the two counts are directly comparable as a trend, not as a 1:1 host match.
host.services.endpoints.http.body:"3aC1db3gGU08O4UHbAt4J"
web.endpoints.http.body:"3aC1db3gGU08O4UHbAt4J"
Host query and web-property query cross-checked against each other (19 hosts; 20 web-property rows, reconciled to 19 distinct IPs once a dual-port host is deduplicated).
For a handful of hosts, I additionally pulled per-service first-seen/last-seen history from Hunt.io to check deployment timing against the disclosure date, verified independently as per-service rather than per-host, since the live dashboard summary and the detailed port-history view did not always agree with each other on the same host.
Findings
The backdoored-console count fell by roughly 59% in four days: 46 at BadByte's 4 Nov census to 19 at this one on 8 Nov. The composition changed as much as the count:
| BadByte, 4 Nov | This census, 8 Nov | |
|---|---|---|
| Backdoored consoles live | 46 | 19 |
| Cloudflare-fronted | 4 | 0 |
| ColoCrossing (HostPapa) cluster | 17 | 1 |
| Luogelang (Hong Kong) cluster | 7 | 0 |
The single-owner fleet in ColoCrossing's address space (15 hosts that BadByte traced to the same SOCKS5-proxy operator) collapsed from 17 to 1. Luogelang's 7-host VPS-template cluster and all 4 Cloudflare-fronted properties disappeared entirely. Both look like straightforward cleanup: a provider or operator pulling backdoored installs once the compromise was public.
One new cluster did not fit that pattern. Four hosts on a single /24, operated by NetLab Global, appeared after the disclosure date, and they're not running on the default port 8787 that the project's own compose file uses. Everything else in the census sits on 8787. A fleet of four, in one address range, moved off the fingerprinted port, is a deliberate choice, not a default deployment. Spot-checking one of these hosts' service history independently showed its entire stack (SSH, HTTP, and the alternate ARTEX port) first appeared two days after BadByte's report went public.
A second host checked the opposite way: one of the surviving consoles' ARTEX service history traces back to late August 2026, roughly two months before the backdoor was even disclosed. That operator has been running a compromised console since well before anyone, including them, likely knew it was compromised.
Read together, these two data points separate two different populations hiding inside the same "still backdoored" count: operators who haven't noticed yet, and at least one operator who kept deploying after the disclosure, on infrastructure deliberately changed to avoid the known fingerprint.
Limitations
This is a count-level comparison, not a host-matched one. BadByte explicitly does not publish the IP addresses behind their 46, in their words, "some of them belong to curious people with no malicious intent, and the backdoor is still open on most of them." That is the right call, and it means there is no way to confirm which specific hosts from their count persisted, changed owners, or went dark; only that the total fell. For the same reason, this report does not publish the current census's live host addresses either; several were still actively exploitable as of this writing, and a public list of working, unauthenticated root-shell targets is not something to publish while they're unpatched.
Both censuses are also a floor, not a ceiling. BadByte separately identified roughly 590 consoles running an unversioned build from the same compromised source tree that cannot be distinguished from a clean self-build without exploiting the backdoor, which is both illegal and not something either investigation did. Hosts behind NAT, VPNs, or firewalls are invisible to internet-wide scanning entirely.
Why this matters
Public disclosure and a registry takedown measurably worked here: a 59% reduction in four days is a real result, and it's the kind of outcome internet-scan data can actually demonstrate rather than assume. But the population that remains isn't uniform. Some of it is simple lag: operators who haven't rebuilt yet. Some of it, based on the NetLab Global cluster's timing and port choice, looks like continued use by someone who knew exactly what they were running.
A disclosure's effectiveness isn't fully captured by "did the number go down." It also matters who's still there once it has.
Detection
If you're running ARTEX from a Docker image, check your own install against the indicators above, or use BadByte's own detection guidance and YARA rules, which this census did not need to reproduce. The short version: the official client never sends a captcha field to the login endpoint. Any request that does (or any login response that comes back with shell output instead of an auth result) is a confirmed hit.
One addition of my own, for hosts where process telemetry is easier to get than HTTP request bodies. The implant runs its command through sh -c from the console binary, so a shell started by that binary is the thing to look at:
title: Shell Spawned by ARTEX Console Binary
id: c4400b2d-efcb-4f71-998e-5a4866fcb638
status: experimental
description: >
The ARTEX console binary starting a shell with -c. This is the execution
path of the backdoored Docker build described in this report.
references:
- https://badbyte.io/artex-backdoor-docker-image/
author: Dheeraj Maram
date: 2026-11-09
tags:
- attack.execution
- attack.t1059.004
logsource:
category: process_creation
product: linux
detection:
selection:
ParentImage|endswith: '/app/artex'
CommandLine|startswith: 'sh -c '
condition: selection
falsepositives:
- The console starting tools through a shell by design. Baseline a clean build first.
level: medium
This rule is experimental. I wrote it from the published behaviour and have not run it against a live backdoored instance, which would mean exploiting one. Treat a match as a reason to check the image hash above, not as proof on its own.
MITRE ATT&CK mapping
| Technique | Tactic | How it shows up |
|---|---|---|
| T1195.002 | Initial Access | A trojanized prebuilt Docker image replaced the archived upstream image after the original project went closed-source, carrying a root-shell backdoor into every install that followed the README's recommended path |
| T1190 | Initial Access | The backdoor triggers on a single unauthenticated POST to the console's own login endpoint; no account or valid credentials are needed |
| T1059.004 | Execution | The implant runs the attacker-supplied command via sh -c and returns the output in the HTTP response |
Indicators
Shown defanged. Switch to refanged before pasting into a tool; STIX 2.1 output is always refanged.
| Type | Value | Context | Copy |
|---|---|---|---|
| File Hash (SHA-256) | d5070f52262effe216b840711343bff8f2787f39208d887d8dfd9c673f8d1573 | Backdoored /app/artex binary inside the trojanized image | |
| UI Build ID | 3aC1db3gGU08O4UHbAt4J | Fixed UI build ID served by every backdoored instance on every page; the fingerprint this census and BadByte's both query on | |
| Docker Image | tomicdd/artex:latest | Removed from Docker Hub 5 Nov 2026 following BadByte's report; 7,913 to 9,131 pulls before removal | |
| Docker Image | congdonglinux/artex:latest | Byte-identical re-push of the same backdoored image; removed the same day | |
| Go Symbol | github.com/Autumn-27/artex/server.verifyCaptcha | The added function; a direct, unauthenticated shell call reachable from the login handler before credentials are checked |
Report: ARTEX Backdoor: An Independent Census of Live Backdoored Consoles