ARTEX Docker Image Backdoor: Root Shell in Login, What to Do
TL;DR · the short version
Researcher Badbyte reported on 4 November 2026 that a Docker Hub image of ARTEX, the open-source AI penetration-testing framework, carried a backdoor in its login handler that gives an unauthenticated caller a root command shell. The image (tomicdd/artex, with an identical congdonglinux/artex) was pushed on 8 October 2026, hours after the ARTEX author said the project would go closed source, and a GitHub copy switched its compose file to pull it. Badbyte says 62 internet-facing consoles served the backdoored build between 2 and 4 November, out of 1,645 ARTEX consoles it saw live, and that Docker removed both images on 5 November at its request. The Hacker News reported the story on 5 November. No source names who planted it or says data was taken from a specific console. The backdoor does not phone home, but data volumes survive a re-pull, so any host that ran it should be treated as compromised.
A copy of ARTEX, the open-source AI penetration-testing framework, was published on Docker Hub with a root shell hidden in its login handler. Researcher Badbyte reported on 4 November 2026 that the image, pushed on 8 October, lets anyone who can reach the console run commands as root without a password, and that it saw 62 internet-facing consoles running the backdoored build between 2 and 4 November. Docker removed the two images on 5 November, Badbyte says. The backdoor is confirmed by Badbyte's analysis; who planted it, and whether anyone used it against a specific console, is not known.
Who this is for: CISOs, red-team and pentest leads, and anyone whose staff or suppliers run open-source offensive tooling in containers.
We read Badbyte's write-up and The Hacker News report, and queried Docker Hub's public API for the two image names, on 17 November 2026. The GitHub copy and its issue thread were deleted, so we could not read them. We did not run ARTEX or any of these images. PlayCISO is not affiliated with Badbyte, ARTEX or any project named. This post gives no trigger details, request formats, console fingerprints, hosts or addresses.
Confirmed vs not yet known
Badbyte says (primary source): two Docker Hub images, tomicdd/artex:latest and congdonglinux/artex:latest, share a modified ARTEX binary whose login handler runs a caller-supplied command as root before any credential check and returns the output. The image was pushed on 8 October 2026; a GitHub copy switched its compose file to it within the hour. 62 consoles served it between 2 and 4 November (a floor); 1,645 ARTEX consoles of any build were live on 4 November. The implant does not phone home. Docker removed both images on 5 November at Badbyte's request.
We checked: Docker Hub's public API returned "not found" for both repositories on 17 November 2026, which fits the removal.
Reported by The Hacker News only: the split of the 62 consoles (26 in the US, 17 in Hong Kong, 7 in China) and the GitHub copy's 2,449 stars and more than 600 forks. Badbyte shows a by-country chart but we could not read its values.
Not yet known: who planted the backdoor, whether anyone used it against a specific console, what was taken from any console, and how many of the pulls became running installs. No source says.
What Badbyte found
ARTEX is one of the agentic offensive frameworks Google's threat intelligence group said attackers had used in real intrusions (see our GTIG post). Per Badbyte, the ARTEX author posted on the morning of 8 October 2026 (UTC) that the tool would go closed source and get no more releases, and the repository came down. That evening a new Docker Hub account pushed an ARTEX image, and in under an hour a GitHub copy changed one line of its compose file to pull that image instead of the official one. People looking for the now-missing project found the copy, which Badbyte says gained about 1,100 stars on 12-13 October, many of which later disappeared.
The change sits in the login handler. Badbyte says the trigger looks like an ordinary failed login from the network, needs no account, and runs the supplied command as root, returning the output before any password is checked. It says the implant adds no new hosts or URLs and never calls home: it waits. In Badbyte's words the change "never shows up in a code review" of the visible project, because it lives in the shipped binary. We do not publish the trigger or the request format.
Why it matters. A pentest console holds other people's secrets. Badbyte lists what a compromised one exposes: database credentials, model and search API keys, captured traffic, notification tokens, the team's unpatched findings and its network position. The attacker can also change the agent's prompts, skills, tools and MCP server definitions, so later runs do something the operator did not ask for. And because data written to the console's volumes survives a re-pull, pulling a clean image over a compromised install does not undo it.
The numbers, read carefully. Badbyte counted 7,946 pulls across the two images by 3 November and 9,131 and 45 by the time Docker removed them; The Hacker News reported "about 9,000". Pulls are not installs. The 62 consoles are internet-facing instances Badbyte saw serving the backdoored build, and it calls that a floor, since consoles kept off the internet would not show up. It does not publish their addresses, saying many belong to people who do not know they are compromised.
Useful next, if this was relevant to you
Timeline
| Date (2026) | Event | Source |
|---|---|---|
| 8 October, morning UTC | ARTEX author says the tool goes closed source; repository taken down | Badbyte |
| 8 October, evening UTC | Docker Hub account tomicdd created and an ARTEX image pushed; a GitHub copy switches its compose file to it within the hour | Badbyte |
| 12-13 October | The GitHub copy gains about 1,100 stars | Badbyte |
| 20 October | First backdoored console seen; congdonglinux/artex registered | Badbyte |
| 23-24 October | A GitHub issue raising a backdoor suspicion is opened, then closed by the copy's owner | Badbyte |
| 30 October | Google GTIG and Mandiant publish "ARTEX and Friends" | GTIG |
| 2-4 November | 62 consoles seen serving the backdoored build; at 09:55 UTC on 4 November, 46 still were | Badbyte |
| 3 November | 7,946 pulls across the two images | Badbyte |
| 4 November | Badbyte publishes its write-up; 1,645 ARTEX consoles live | Badbyte |
| 5 November | Docker removes both images at Badbyte's request; the tomicdd account is deleted. The Hacker News reports the story | Badbyte (update), The Hacker News |
| 17 November | Docker Hub's public API returns "not found" for both repositories | PlayCISO check |
Indicators: where to get them
Badbyte publishes an indicator section for defenders: the two image names (tomicdd/artex:latest and congdonglinux/artex:latest), the image index, platform manifest, layer and binary SHA-256 digests, the GitHub repository and the commit that swapped the compose image, the Docker Hub and GitHub account names, and two YARA rules. Take them from Badbyte's page rather than from a copy, and match on digests, not names: Badbyte says about 34 Docker Hub namespaces publish "artex" images and only two carried the backdoor. It deliberately publishes no host addresses.
What security leaders should do this week
- Find every ARTEX, and every fork. Ask red-team, pentest and research staff, and suppliers who test for you, whether they run ARTEX or a fork such as AntiFox. Search container hosts, compose files and registries for any image named "artex" and note where each compose file pulls from.
- Compare digests, not names. Check every image and the binary inside it against Badbyte's published digests and YARA rules. A clean name is not proof: a fork may have renamed the image, and mirrors may still serve a cached copy.
- Treat a host that ran the backdoored image as compromised. Stop it and isolate it. Keep its volumes for forensics but do not reuse them. Rotate everything it could reach: model provider and search API keys, database passwords, webhook, bot and mail tokens, MCP credentials and any target credentials it stored. Review agent prompts, skills, tools and MCP definitions for changes. If it held client or third-party data, think about telling them. Work through it with the AI incident response checklist and the incident response hub.
- Hunt on systems you own. Badbyte suggests alerting when the ARTEX process spawns a shell and flagging login requests that carry fields the official client never sends. Look back over your logs to 8 October. Apply checks only to systems you own or are authorised to test.
- Never pull unofficial images. Run offensive tools only from an upstream release you have verified, built from source you have read, pinned by digest, and with signatures checked where the project signs. When a project goes closed or disappears, freeze on your last verified copy rather than reaching for a mirror.
- Keep consoles off the internet. Bind them to localhost or a VPN, never as root, never on an open port. Check hosts you own with the Agent Exposure Scanner, and score your wider readiness with the Agentic Attack Readiness Scorecard.
- Write it into policy. Require approval and an owner for any offensive tool, an approved-source list for containers, and an inventory of where those tools run.
Where this fits
This is the second ARTEX lesson in a month. The first was that attackers use agentic frameworks at machine speed (GTIG); this one is that the people who run them are a target too, because a pentest console concentrates credentials and findings. The same week, India's CERT-In published an advisory naming ARTEX among the tools used against financial firms and an AI security auditing guideline; our CERT-In Crafty Falcon post covers that, and no source links that actor to the backdoored image. For exposed consoles in general, see what to hunt.
Frequently asked questions
What is the ARTEX Docker image backdoor?
Badbyte reported on 4 November 2026 that a Docker Hub image of ARTEX, the open-source AI pentest framework, had been modified so that its login handler runs a command sent by an unauthenticated caller as root, before any password check, and returns the output. The image was pushed to Docker Hub on 8 October 2026, hours after the ARTEX author said the project would go closed source. It was not the official ARTEX image.
Which images are affected?
Badbyte lists two images as indicators: tomicdd/artex:latest and congdonglinux/artex:latest, which it says share the same digests. It also names a GitHub copy, trhhubsd/ARTEX (earlier ARTEXX), whose compose file was changed to pull the first image. Badbyte says about 34 Docker Hub namespaces publish "artex" images and only those two carry the backdoor. Badbyte publishes the image, layer and binary digests on its page; check against those rather than names alone.
Is the backdoored image still on Docker Hub?
Badbyte says Docker removed both images on 5 November 2026 at its request, and that the tomicdd account was deleted. Docker Hub's public API returned "not found" for both repositories when we checked on 17 November. Removal does not clean copies already pulled, images cached by third-party mirrors, or forks that renamed the image, which Badbyte says can still be exposed.
How many servers were exposed?
Badbyte says it saw 62 internet-facing consoles serving the backdoored build between 2 and 4 November 2026, and calls that a floor. At 09:55 UTC on 4 November, 46 were still backdoored, 4 had been rebuilt on other images and 12 were unreachable. It counted 1,645 live ARTEX consoles of any build that day. These are scan counts of exposed consoles, not a count of confirmed intrusions.
How many times was the image pulled?
Badbyte counted 7,946 pulls across the two images by 3 November and final counts of 9,131 and 45 when Docker removed them on 5 November. The Hacker News reported about 9,000 downloads. Pull counts include repeats and automation, so they are not a count of installs.
Did the backdoor steal data or phone home?
Badbyte says it does not phone home: it adds no new hosts or URLs and waits for a request. No source says data was taken from any particular console. The risk Badbyte describes is what an attacker could reach on a console: database credentials, model and search API keys, captured traffic, notification tokens, the team's findings, and the agent's prompts, skills, tools and MCP server settings, which could be altered for later runs.
Who planted the backdoor?
No source names a person or group. Badbyte describes a Docker Hub account created the evening of 8 October and a GitHub account that switched the image within the hour, both since deleted. Attribution beyond those accounts is not known.
Does updating the image remove the backdoor?
Not reliably. Badbyte says data written to the data, skills and database volumes survives a re-pull, so a compromised console's stored secrets and any altered agent settings stay behind. Treat the host and its volumes as compromised and rebuild from a source you have verified.
Is the official ARTEX image or AntiFox backdoored?
Badbyte says official v0.3.26 builds do not match its backdoor rule, and it names only the two images above. Nothing we read says AntiFox, an ARTEX fork, is backdoored. Any fork that took its compose file from the trojanized copy could still point at the bad image, so check where yours pulls from.
How do I check my own ARTEX install?
Badbyte's detection section suggests checking which image and compose source you run, comparing the binary hash inside the container with its published value, watching login requests for fields the official client never sends, alerting when the ARTEX process spawns a shell, and running its two YARA rules. Apply those checks only to systems you own or are authorised to test.
Sources
- Badbyte: write-up on the backdoored ARTEX Docker image, with detection, YARA rules and indicators (4 November 2026, updated 5 November); accessed 17 November 2026
- The Hacker News: CERT-In Flags AI-Driven Attacks as Backdoored ARTEX Exposes 62 Consoles to Root Access (5 November 2026); accessed 17 November 2026
- Google GTIG and Mandiant: ARTEX and Friends: Enter Agentic Offensive Frameworks (30 October 2026)
- CERT-In: Artificial Intelligence (AI) Security Auditing Guidelines (4 November 2026)