AI SecurityAgentic AIThreat ResearchSupply Chain

Somebody Backdoored the AI Hacking Tool, and Microsoft Clocked Its Agents at Five Minutes

November 6, 2026 · Brian Deitch

I've written about ARTEX, the AI-driven hacking framework, a few times now, plus the Key Eater campaign. (Usual disclosure: I work at Zscaler, whose ThreatLabz team was one of the first to publish on ARTEX.) This week brought two new twists, and one of them is almost funny.

Twist One: The Hackers Got Hacked

On November 4, independent researcher BadByte published an analysis of a popular prebuilt ARTEX Docker image on Docker Hub (tomicdd/artex, with an identical copy at congdonglinux/artex). It was backdoored.

Per BadByte, a function called verifyCaptcha in the login handler takes the CAPTCHA field and runs it with sh -c, as root, and sends the output back. Their summary: "It runs whatever command arrives in the login form, before the password is checked." So anyone who knew the trick had an unauthenticated root shell on every console running that image.

The numbers from BadByte:

  • About 7,946 pulls by November 3, before disclosure, climbing past 9,000 afterward.
  • 62 consoles seen serving the backdoored build (BadByte calls that a floor), with 46 still up on the morning of November 4.
  • 1,645 ARTEX consoles live on the internet that day.

Important detail: BadByte says the GitHub source was clean. "The backdoor lives only in this prebuilt image." Docker removed both images on November 5, but any copy already pulled or mirrored is still backdoored.

India's CERT-In also issued an advisory this week about an active campaign it calls "Crafty Falcon" targeting the banking and financial sector with ARTEX and other AI offensive tools.

Twist Two: Five Minutes From GitHub to Key Vault

On November 5, Microsoft published research tying ARTEX to the hunt for exposed secrets. Threat actors used the framework to search public GitHub for OAuth app secrets, app credentials and AI service API keys, starting in August 2026, then used what they found to sign in as Entra service principals from as early as October 28.

The line everyone will quote, because it deserves it: "an AI agent progressed from searching for the client secrets in GitHub to exfiltrating data from Azure Key Vault in five minutes."

Microsoft's main IP, 199.19.77.84, is the same one Todyl flagged in the Key Eater research I covered on Wednesday. So that campaign very likely has a tool name attached now.

Why This Matters

Two lessons, one for everyone and one for the people running these tools.

For everyone: five minutes is less time than it takes most teams to triage the alert. A leaked secret used to be a "rotate it this sprint" problem. When an agent is watching GitHub, it's a "rotated already, or it's too late" problem. As Microsoft puts it, "Publicly exposed credentials remain usable until revoked or rotated."

For red teams and researchers: if you're pulling a prebuilt image of an offensive tool from a random Docker Hub account and exposing its console to the internet, you're running untrusted code with root and a pile of LLM API keys sitting in plaintext. That's a supply chain attack on you. Offensive tools are software too.

What To Do

  • Running ARTEX? If you pulled either image, assume the host is compromised. Rebuild from source you've read, keep the console off the internet, and rotate every secret it held (BadByte notes the database password and LLM keys are stored in plaintext). Check login request bodies for a captcha field and alert on sh -c spawned by /app/artex.
  • Turn on secret scanning for public repos, and treat a detected leak as an incident: revoke first, investigate second.
  • Move workloads to managed identities where you can, so there's no client secret to leak.
  • Follow Microsoft's response steps if you see the IOCs: block 199.19.77.84 with Conditional Access, revoke sessions and Azure DevOps PATs, remove app secrets and certificates, and rotate everything in any Key Vault those apps could read.
  • Least privilege for apps. A service principal that can read every vault turns one leaked secret into all of them.

Bottom Line

Attackers are using AI agents to turn a leaked secret into stolen data in minutes, and at least some of those attackers were running a tool that was quietly handing out root on their own boxes. Rotate fast, scope tight, and don't run code you didn't check, even when you're the one doing the hacking.

Sources: BadByte on the ARTEX backdoor, Microsoft Security on agentic AI and exposed secrets, The Hacker News on the ARTEX image and CERT-In, CERT-In advisory CIAD-2026-0058, and Todyl on stolen app credentials

artexdocker-hubbadbytemicrosoftcert-inagentic-aiexposed-secretsgithubentra-idazure-key-vaultsupply-chain
All postsDiscuss on LinkedIn