The Backdoor Was in the Login Field

November 9, 2026 · SPR{K3 Research

On November 5, independent researcher Badbyte sent The Hacker News a short, ugly finding. A Docker Hub image named tomicdd/artex — pulled about nine thousand times — was running on sixty-two internet-facing ARTEX consoles between November 2 and 4. The image is a trojanized build of the ARTEX offensive-AI pentest framework; the Cloud Security Alliance relayed the technical details the next day in a research note that attributes them to the press reporting.

The backdoor is not hidden in some exotic corner of the code. It is in the login handler, in a function named verifyCaptcha().

If a login request's CAPTCHA field starts with a secret marker, the function strips the marker, passes the rest of the value to sh -c, runs it as root, and puts the output in the HTTP response. Before any credential check.

One HTTP request to the login page of any ARTEX deployment running this image, and you have a root shell.

Where it came from

The image was pushed around October 8. The earliest internet-facing serving was October 20. By November 4, 62 of the 1,645 hosts serving some version of ARTEX were serving the trojanized one — 26 in the US, 17 in Hong Kong, 7 in China, 12 elsewhere.

Delivery was quiet. The attacker published a GitHub fork, trhhubsd/ARTEX, that changed exactly one line in docker-compose.yml: the image: field, repointed to the trojanized Docker Hub namespace. The application source tree in the fork was visually unchanged.

An operator who cloned the fork, read the source, and ran docker compose up deployed a container with pre-auth root RCE wired into its front door. Reading the source would have shown nothing. The payload was in the image.

Why verifyCaptcha()

Badbyte put the point crisply in the writeup: the backdoor code "never shows up in a code review."

Not because it is hidden. Because a function called verifyCaptcha() is exactly what a reviewer skimming an authentication handler expects to find. Anti-abuse logic lives there. The eye slides past.

Three choices stack on top of that:

These are the same choices every good backdoor makes. What is new is the surface they are on. The vulnerable workload is a privileged offensive-AI console — configured with operator credentials, tool chains, API keys, and target lists. Someone who gets root there inherits the attacker's whole operating position.

The attack surface is the attacker

ARTEX is the framework a researcher's runaway sandbox used to breach WorkOS in late October, which the same reporting ties to the Vapor Panda ~200-organization campaign, the DIVD disclosure, and intrusions against South Korean banks.

This new finding is not about ARTEX being used to attack someone. It is about ARTEX being attacked. The user base is the people who deploy an offensive-AI framework to run operations against other targets. The trojanized image gives a second attacker root on the first attacker's console.

The supply-chain mechanics are not new. One-line docker-compose.yml diff in a lookalike fork, lookalike namespace on Docker Hub, benign-looking source tree — primitives any decade-old supply-chain attack would recognize. What has changed is the artifact. The thing being substituted is no longer a generic library. It is an AI agent framework running as root with the keys to a live operation.

The pattern

This is the second time in a week a published finding has put the backdoor inside the defender control itself. On November 5 Check Point disclosed a Lockdown Mode bypass in ChatGPT where the sandbox permissions that were supposed to restrict rendered HTML were themselves the escape path. Different product, different layer, same shape — the attack lives where the review expects to see anti-abuse logic.

Where Defend sits

The backdoor is a single request. There is nothing a prompt-layer filter can see about it. There is nothing a source-code review would have caught. The container pulled at runtime is the attack.

The sequence that follows an exploitation hit — a child process spawning from the login handler, shell output flowing back out the HTTP response, outbound connections initiated from inside the container, all from an unauthenticated request — is what a runtime observer watching the workload sees. Defend is built to watch that shape. The individual primitives are legitimate in isolation. The attack is their ordering.

What to do now

The source was unchanged. The image was the attack. The login field was the trigger. The compose file was the delivery.


SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.