What Is ARTEX? The AI Pentest Framework, the Backdoored Docker Image, and How to Respond

November 17, 2026
November 17, 2026

0 min read

Vulnerabilities
AI Security
ARTEX AI Pentest Framework: Backdoored Docker Image, IOCs, and Response Runbook

ARTEX is an open-source, LLM-driven penetration testing framework, published on GitHub on July 26, 2026, that attackers now use to automate intrusions from reconnaissance through data exfiltration. The FBI described ARTEX in FLASH-20261113-001 on November 13, 2026, and reported that some intrusions moved from initial targeting to privileged access in under three hours.

This reference, current as of November 2026, covers what ARTEX (also written ARTEX AI or Artex) is, who uses the framework, the backdoored tomicdd/artex Docker image, the indicators to hunt for, and a first-hour runbook. One section answers the question of the week: which recent incidents are linked to ARTEX, and which are not. Security teams running ARTEX for authorized testing get their own section: the tool became a supply chain target the day its author took the repository down.

ARTEX at a glance

ARTEX is an attack tool, not a vulnerability, so no CVE exists and no patch applies.

FieldValue
NameARTEX (ARTEX AI, Artex; commercial edition ARTEX PRO)
TypeOpen-source, multi-agent AI penetration testing framework
Author and licenseAutumn-27 on GitHub, AGPL-3.0
First publishedJuly 26, 2026
Taken closed-sourceOctober 8, 2026, 10:50 UTC, after reported abuse
StackGo server, Next.js console, PostgreSQL, usually run in Docker
Default console port8787
CVENone. ARTEX is a dual-use tool
Exposed consolesMore than 1,600 as of early November 2026, per the FBI
FBI alertFLASH-20261113-001, November 13, 2026
Backdoored imagestomicdd/artex, congdonglinux/artex (removed from Docker Hub November 5, 2026)

ARTEX hands a penetration test to a team of AI agents

ARTEX takes a goal in natural language and has LLM agents plan, execute, and record the test with real tools. A person sets the objective and steers. The agents do the work.

The last open-source release, version 0.3.15, works like this:

  • One program, one console. A Go server, a Next.js web console, and a PostgreSQL database run together, usually in Docker. The supplied Compose file publishes port 8787 on every interface, so on a rented server the console faces the internet at startup.
  • Planner and workers. One agent turns the operator's goal into objectives and scope. A planner breaks objectives into steps. Worker agents, three per task by default, run tools such as nmap, a shell, and an HTTP client, then write back results. A retest agent checks findings a second time.
  • Bring your own model. ARTEX ships no model. Operators connect Anthropic, OpenAI, or any compatible API. CrowdStrike and the FBI report attacker deployments running mostly on DeepSeek models.
  • Soft limits only. The original blocks .gov and .edu targets and holds risky actions for approval. CERT-EU found no hard technical enforcement of scope, and at least one fork advertises removal of the blocklist.

Most ARTEX techniques are old: scanning, password spraying, exploits for known web flaws such as SQL injection, SSRF, and anonymous file upload. The agents add speed and persistence. Microsoft watched one agent send a malformed request to Azure, read the error, and send a corrected request about 28 seconds later.

ARTEX use spread from one criminal ring to state-aligned groups in three months

No single group owns ARTEX. The FBI says a variety of actors have used the framework since at least August 2026, and published reporting names at least six. Barracuda added a confirmed victim report on November 16.

ActorReported byWhat happened
Storm-0940 (China state-aligned)Microsoft, WizUsed ARTEX against hundreds of organizations. Stole the credentials of the ConnectWise Cloud Backup application, then attempted sign-ins to more than 2,250 customer tenants. Some intrusions reached privileged cloud access in under three hours.
VAPOR PANDA (China-linked)CrowdStrikeMore than 200 organizations in 14 countries in 11 days, mostly government and defense.
Chinese-speaking actor, South Korean financeCrowdStrike, CensysCompromised at least nine organizations, including major banks, in late September and early October. Agents deployed web shells and stole credentials. The operator left ARTEX settings and AI session logs in an open directory.
Storm-3358 (Cambodia-based, criminal)MicrosoftFinancial firms in Southeast Asia since October. Agents exploited anonymous file upload and SSRF flaws, then installed remote access tools.
Crafty Falcon (criminal)CERT-In advisory CIAD-2026-0058Indian lenders, fintechs, and crypto exchanges. Enters through test and staging systems, then reuses credentials shared with production.
South Korean criminal ringSeoul Metropolitan PoliceMore than 90 organizations and 8.2 million records. Leader arrested October 15.

What the agents have exploited

ARTEX ships no exploits of its own. The agents use whatever weakness an internet-facing system exposes, and the reporting so far names classes more often than CVEs.

  • Web flaw classes. SentinelOne lists IDOR, command injection, and server-side template injection, next to hardcoded administrator passwords and unauthenticated APIs. Check Point documented command injection and template injection used for remote code execution against iOS signing services, without naming products or CVEs.
  • CVE-2022-2588 and CVE-2025-5054. Operator infrastructure held exploit builds with these Linux kernel CVEs in the file names, per SentinelOne. That shows the builds were present. No report confirms the agents ran them.
  • CVE-2026-75650. AhnLab found a working copy of the public StyleSmuggler exploit for this Magento flaw inside REAPER, a customized ARTEX distribution.

The fastest path in is a credential, not an exploit

ARTEX intrusions lean on exposed credentials more than on software flaws. Microsoft observed an agent move from a GitHub search for a target's secrets to data theft from Azure Key Vault in five minutes. Unit 42 saw campaigns reach cloud access through password reset pages and exposed APIs without exploiting a single vulnerability. GitGuardian found no secret-hunting feature in the ARTEX code: the model decided to search code and git history for credentials unprompted, and found eight of nine planted secrets in lab runs.

No public evidence links ARTEX to the Optimizely attack or the US outages

As of November 17, 2026, no public reporting ties ARTEX to the Optimizely attack or to the outages at JPMorgan Chase, United Airlines, and the NSA. The FBI alert is dated November 13 and landed in the same news cycle, which is why the stories keep getting joined.

IncidentWhat is reportedARTEX link
Barracuda cloud breach (November 1 to 2)An attacker used a cloud access key issued to a third-party vendor and never retired, and reached about 10 cloud accounts over 44 hours.Confirmed by Barracuda's investigation
Optimizely, also called OPTIMJACK (November 16)An attacker exploited a flaw in Optimizely Web Experimentation and injected JavaScript served to visitors of at least 178 organizations, per Unit 42.None reported
JPMorgan Chase outage (November 16)The bank attributed the outage to a software problem in its own systems, per CoinDesk.None reported
United Airlines outage (November 16)United said it found no indication of a cybersecurity incident.None reported
NSA network disruption (November 17)The NSA described temporary degradation during scheduled routine maintenance.None reported

Officials have not linked any of the outages to ARTEX, to one another, or to a cyberattack. ARTEX has been used against banks, in South Korea, Southeast Asia, and India. Those cases are separate from the US outage.

The Barracuda case is the pattern to learn from: a vendor key outlived the project, and an agent found the gap.

The tomicdd/artex Docker image gave anyone a root shell before login

A backdoored ARTEX image on Docker Hub, tomicdd/artex, runs any command sent in the login form as root, before the password is checked. BadByte published the analysis on November 4, 2026, and Docker removed the image and an identical copy, congdonglinux/artex, on November 5.

The sequence took under ten hours on October 8, 2026 (all times UTC):

  1. 10:50. The author announces ARTEX is going closed-source and removes the repository.
  2. 18:50. A new Docker Hub account, tomicdd, is created. The account pushes tomicdd/artex:latest at 19:41.
  3. 20:33. A GitHub copy, trhhubsd/ARTEX, changes one line of docker-compose.yml in commit a4f4d38: image: autumn27/artex becomes image: tomicdd/artex. The README presents the copy as the maintained successor.
  4. October 12 to 13. The repository gains 1,098 stars in 48 hours. About 1,100 stars later disappear.

The root cause is a trust gap between source and artifact. The repository source matched upstream, so a code review found nothing. The backdoor lived only in the prebuilt binary at /app/artex: a 160-byte function, server.verifyCaptcha, that passes a captcha field from POST /api/auth/login to sh -c and returns the output. The implant never calls out. From the network, the trigger looks like a failed login.

MeasureValueSource
Pulls before disclosure (by November 3)7,946 (tomicdd/artex 7,913, congdonglinux/artex 33)BadByte
Final pulls of tomicdd/artex9,131BadByte
Internet-facing consoles serving the backdoored build62 (November 2 to 4)BadByte
GitHub repositories still pointing at the image637BadByte
Live ARTEX consoles of any build1,645 (November 4)BadByte

One shell in the container reaches everything ARTEX stores in plaintext: LLM provider keys in ANTHROPIC_API_KEY and OPENAI_API_KEY, the database password in ARTEX_PG_DSN, the session signing key at /app/jwt.key, credentials harvested from engagement targets, and unpatched findings. Updating does not remove the backdoor. The working update path, update.sh, runs docker compose pull against the same image.

Public reporting does not attribute the image to Crafty Falcon. CERT-In names ARTEX among that group's tooling, and the two stories broke the same day, but no source links the operator.

Indicators of Compromise (IOCs) and behavioral signals

Hunt in three places: traffic against your applications, your identity logs, and any ARTEX install your own team runs. ARTEX infrastructure changes fast, so treat IP lists in the primary reports as a starting point, and note that authorized testers also run ARTEX.

Backdoored image and binary

TypeIndicator
Docker imagetomicdd/artex:latest
Docker imagecongdonglinux/artex:latest
Manifest digest (linux/amd64)sha256:cb85a7206290dc1813e3203e044741054b26c7a7c557c7d9d6d91af20aeb232b
SHA-256 of /app/artexd5070f52262effe216b840711343bff8f2787f39208d887d8dfd9c673f8d1573
UI build ID served on every page3aC1db3gGU08O4UHbAt4J
Symbol in binarygithub.com/Autumn-27/artex/server.verifyCaptcha
Module version stringv0.3.16-0.20261008134550-5c0dbd4bec94+dirty
Request patternPOST /api/auth/login with a captcha key in the JSON body
Host signalsh -c spawned by /app/artex

ARTEX activity against your systems

TypeIndicatorSource
HTTP user agentsartex-enrich/1.0, norma/0.4, Mozilla/5.0 (compatible; ArtEx/1.0)WorkOS, Wiz
Request rhythmBursts of 1 to 14 requests about a second apart, then a 25 to 31 second pauseUnit 42
Test inputsevil[.]com, test@example[.]comUnit 42
Entra ID sign-in errorsAADSTS7000215 (wrong client secret), AADSTS7000222 (expired client secret)Microsoft
Console fingerprintPort 8787, page title starting ARTEXCensys, SOCRadar
Internal signalAI model API traffic from servers with no business reason to send anyCERT-In

A failed application sign-in followed by a success from the same source is the urgent case. Either Entra error means someone holds a credential for your application.

Immediate response: the first-hour runbook

Start with your own ARTEX installs, then credentials, then logs. An install of the backdoored image is a confirmed compromise. Everything else is a hunt.

  1. Find every ARTEX install. Run docker compose config | grep -nE 'image: *(tomicdd|congdonglinux)/artex' and docker image inspect --format '{{json .RepoDigests}}' tomicdd/artex congdonglinux/artex on any host where a security or research team tests tooling.
  2. Verify the binary. Run docker exec <container> sha256sum /app/artex and compare against the hash in the IOC table. Check your own console for the build ID with curl -s http://127.0.0.1:8787/ | grep -c 3aC1db3gGU08O4UHbAt4J.
  3. Isolate and rebuild any host that ran either image. Removal from Docker Hub does not clean existing installs or mirror caches. Data in ./data, ./skills, and the database volume survives a re-pull, so rebuild from a trusted source.
  4. Rotate everything the console held. LLM provider keys, the database password, the JWT signing key, MCP server tokens, notification webhooks, and every credential harvested from engagement targets. Tell the owners of those targets.
  5. Search web, proxy, and API gateway logs for the user agents artex-enrich/1.0 and norma/0.4, and for the burst-then-pause request rhythm.
  6. Review application sign-ins in Entra ID. Look for AADSTS7000215 and AADSTS7000222, and for one IP address signing in as several applications.
  7. Scan repositories for secrets, including full git history, and rotate every hit. Microsoft's guidance is direct: removing the original disclosure does not remediate the exposure.
  8. Report confirmed ARTEX activity to the FBI or IC3, as the FLASH requests.

A scanner shows what is present. Runtime shows what executed.

Every static check passed in the tomicdd/artex case, and the only evidence of the backdoor was a process spawning a shell from a login handler. The same gap decides how an ARTEX intrusion goes on the defender's side: agents test every exposed path at machine speed, so the question that matters is which weakness is real in production, not how many findings exist.

Kodem's layers map to the specific failures in this story:

  • Exposed secrets. Kodem secret scanning finds exposed secrets in code and git history, with policies that can block a merge. Git history is where ARTEX agents found credentials in GitGuardian's lab runs.
  • Container images. Kodem scans Docker and OCI images in the registry and at runtime, detects malware in scanned images, and separates third-party vendor images from customer-built ones as part of securing the SDLC. A backdoor made of one small function with no network indicators is a hard target for any image scanner, which is why the next layer matters.
  • Runtime truth. Runtime intelligence shows which packages a running process has loaded and which vulnerable functions are invoked. The fix queue starts with code that executes.
  • Exploitation in the running application. ARTEX agents exploit web flaws such as SQL injection, SSRF, and file upload. Kodem Application Detection and Response (ADR) detects exploit behavior at the moment of initiation, using behavioral baselines and call-stack context instead of signatures. Call-stack context separates a process spawn the application intends from one reached through an injection path.

Two boundaries, stated plainly. Kodem is not an autonomous pentest tool and does not do what ARTEX does. ARTEX explores broadly to discover. Kodem investigates one specific risk, validates whether that risk is exploitable in your application, and verifies that the response removed the risk, for supported risk classes. Kodem also works at the application layer and does not replace identity controls, WAF, or EDR. Credential abuse against Entra ID belongs with those controls.

Hardening against the next agentic framework

Remove the easy credentials and keep agent tooling off the internet. Blocking one tool's features will not help, since about 45 similar open-source frameworks exist, per GitGuardian.

  1. Replace long-lived cloud keys in CI with short-lived OIDC tokens. Set an expiry on every personal access token and prefer fine-grained tokens.
  2. Move Entra ID applications from client secrets to managed identities or certificates where possible.
  3. Share no credentials, tokens, or API keys between test and production. This is CERT-In's first recommendation on Crafty Falcon.
  4. Retire vendor keys when the project ends. The Barracuda breach ran on a key issued to a vendor for work that had finished.
  5. Pin container images by digest from a source you verified. A pin protects only if the pinned version was safe. Build security tooling from source you have read.
  6. Keep agent consoles off the internet. Bind port 8787 to localhost and reach the console over a VPN or SSH tunnel.
  7. Enforce agent scope with network controls, not only the tool's own settings. The WorkOS incident began when a researcher's agent left its test environment.
  8. Alert on AI model API traffic from servers with no development purpose.

ARTEX is the first widely abused framework of its kind, not the last

Seeing ARTEX in an intrusion no longer tells a defender who is behind the activity or what they want. The tool is a commodity, the techniques are familiar, and the difference is tempo.

  • The takedown did not contain the tool. Forks, localized versions, and ARTEX PRO remain available, per the FBI.
  • Offensive tooling is now a supply chain target. A console holds its targets' credentials, findings, and network position, which makes the operator worth compromising.
  • Speed moves the advantage to whoever has context. A team that knows which weaknesses are real in production can close them in the hours an agent needs to find them.

One prediction to check later: at least one more trojanized distribution of an agentic offensive framework will be reported before mid-2027.

Frequently Asked Questions

What is ARTEX?

ARTEX is an open-source, multi-agent AI penetration testing framework written in Go and published on GitHub on July 26, 2026. LLM agents plan a test, run real tools, and record findings with little human input. Reporting also calls the tool ARTEX AI or Artex.

Is ARTEX malware?

No. ARTEX is a dual-use offensive security framework built for authorized testing. Attackers adopted the same tool, which is why the FBI issued FLASH-20261113-001.

Is there a CVE or a patch for ARTEX?

No CVE exists and no patch applies, since ARTEX is an attack tool and not a vulnerability. The defense is to close what the agents look for: exposed credentials, internet-facing internal tools, and unpatched web flaws.

Who is using ARTEX in attacks?

Published reporting names Storm-0940, VAPOR PANDA, Storm-3358, Crafty Falcon, a South Korean criminal ring, and a Chinese-speaking actor targeting South Korean finance. The FBI dates malicious use to at least August 2026.

Is ARTEX behind the Optimizely attack or the US outages?

No public reporting links ARTEX to the Optimizely attack of November 16, 2026, or to the outages at JPMorgan Chase, United Airlines, and the NSA. Officials have not tied any of the outages to ARTEX or to a cyberattack. Barracuda is the recent incident with a confirmed ARTEX link.

Was ARTEX backdoored?

The ARTEX source was not. Two prebuilt Docker Hub images were: tomicdd/artex and congdonglinux/artex. Both ran commands from the login form as root before authentication, and Docker removed both on November 5, 2026.

How do I know if I ran the backdoored ARTEX image?

Check Compose files and image digests for tomicdd/artex or congdonglinux/artex, then compare the SHA-256 of /app/artex against d5070f52262effe216b840711343bff8f2787f39208d887d8dfd9c673f8d1573. A match means the host is compromised. Rebuild and rotate every stored credential.

How do I detect ARTEX activity against my applications?

Search logs for the user agents artex-enrich/1.0 and norma/0.4, bursts of requests followed by 25 to 31 second pauses, and Entra ID errors AADSTS7000215 and AADSTS7000222. Treat each match as a lead, since authorized testers also run ARTEX.

Is ARTEX still available after the takedown?

Yes. The author closed the source on October 8, 2026, but the FBI says ARTEX remains available through forks, archived copies, localized versions, and a commercial edition called ARTEX PRO.

The bottom line

ARTEX did not invent a technique. ARTEX removed the time defenders used to have. The response is not another scanner producing more findings. The response is knowing, with runtime evidence, which weaknesses are real, and closing those first.

References

  1. FBI, FLASH-20261113-001: Threat Overview of ARTEX Autonomous Penetration Testing Framework, November 13, 2026
  2. Microsoft Threat Intelligence, Storm-0940 uses ARTEX agentic AI in ConnectWise supply chain compromise, October 21, 2026
  3. BadByte, ARTEX Backdoor: 7,946 Pulls of a Trojanized Docker Image Handing Out Root Shells, November 4, 2026
  4. CERT-EU, From Penetration Testing to Cyber Threat: Beware of ARTEX
  5. StepSecurity, ARTEX: Attackers Adopted This AI Pentesting Tool in Three Months, November 12, 2026
  6. BleepingComputer, Barracuda links cloud breach to ARTEX AI-powered cyberattack, November 16, 2026
  7. Unit 42, Borrowed Code, Real Consequences: What OPTIMJACK Teaches Us, November 2026
  8. Kodem, Application Detection and Response (ADR)
  9. Kodem, Runtime Intelligence
Table of contents

Related blogs

Java4Shell: CVE-2026-114938 Explained

Java4Shell: CVE-2026-114938 Explained

Java4Shell (CVE-2026-114938) is a pre-auth OpenJDK RCE in Java's Base64 decoder. Get affected Java 17, 21, 25 builds, the Java4Shell patch and mitigation flag.

November 18, 2026

Kodem Now Connects to Microsoft Copilot Studio

Kodem Now Connects to Microsoft Copilot Studio

Kodem now connects to Microsoft Copilot Studio, in beta. Inventory your agents and check each planned tool call against your AI policies.

November 16, 2026

5

When Everything Becomes Open-Source

When Everything Becomes Open-Source

Our research shows that an AI agent can reconstruct a working copy of an application when given the original software's binary or package.

November 10, 2026

13

Stop the waste.
Protect your environment with Kodem.

A Primer on Runtime Intelligence

See how Kodem reads what actually loads and executes in your running applications, and why that changes which findings matter.

See the Kodem platform

Kodem shows which vulnerabilities actually load and execute in your running applications, so your team works the risk that is real.

The State of the Application Security Workflow

This report aims to equip readers with actionable insights that can help future-proof their security programs. Kodem, the publisher of this report, purpose built a platform that bridges these gaps by unifying shift-left strategies with runtime monitoring and protection.

3D book mockup of Kodem's State of the Application Security Workflow 2025 report

Get real-time insights across the full stack…code, containers, OS, and memory

Watch how Kodem’s runtime security platform detects and blocks attacks before they cause damage. No guesswork. Just precise, automated protection.

Kodem issues list with a magnified view of insight icons: runtime, ingress, and exploitability
Combined author
Kodem Security Research Team
Publish date

0 min read

Combined author
Mahesh Babu
Publish date

0 min read

Vulnerabilities

AI Security