top of page

Barracuda Gave an AI Vendor a Cloud Key and Never Took It Back. An ARTEX Operator Found It and Walked Through 10 Cloud Accounts in Under Two Days. Go Find Your Forgotten Vendor Keys

Writer: Patrick Duggan
Patrick Duggan
1 hour ago
6 min read

List every cloud access key you have handed to an outside company, and kill the ones whose project is over. That is the whole lesson of Barracuda's breach, in Barracuda's own words: it gave a cloud access key to an AI vendor for a past project, did not retire it when the work ended, and wrote "That is on us." If you use Barracuda Managed XDR Cloud or Monitored Endpoint and Barracuda contacted you directly, revoke and recreate your integration credentials and review your connected systems' activity logs from November 1, 2026. If Barracuda did not contact you, it says no action is needed.


This is the fourth time we have written about ARTEX, the open-source AI penetration testing framework that criminals and state groups have been running since the summer, and the first time a victim has named itself and published how it happened. Our earlier posts: CrowdStrike's VAPOR PANDA report, the backdoored Docker image and the FBI's FLASH.



What happened, per Barracuda


Barracuda published a Trust Center advisory on November 10, updated November 13, and a post by its CISO, Arve Kjoelen, dated November 11. BleepingComputer reported it Monday.


The key. Barracuda had given a cloud access key to an unnamed third-party AI company. The vendor's development environment was compromised in October, and the key was sitting in a configuration file in the vendor's GitHub repository. Barracuda does not say whether that repository was public or private, or which cloud the key belonged to.


The intrusion, with Barracuda's times in UTC:


  1. November 1, 04:33. The first call with the key. Within minutes the attacker had mapped what the key could reach and read stored secrets, including AWS access keys and Azure service principal secrets.

  2. The next day or so. Each secret opened the next system: about ten cloud accounts, a data analytics system, a source-code access token, then Kubernetes clusters. The attacker decrypted infrastructure-as-code state files and pulled credentials from server metadata services and Kubernetes service accounts.

  3. Early November 2. In under four hours, it pushed short-lived branches to several repositories to trigger a build workflow that printed build secrets into the job log, encoded so the CI system's masking would not hide them.

  4. On servers where it got a shell, it searched for environment variables named WALLET, MNEMONIC, SEED and PRIV, Ethereum keystore files, extended private keys, seed phrase files and wallet-named Kubernetes secrets. Barracuda says there was no wallet material there.

  5. For persistence it deployed a container that mounted the underlying server's file system. The mount was read-only, but that was enough to take the server's credentials. Its attempt to create a new cloud administrator identity was blocked.

  6. November 2, 23:31. The last recorded activity, about 43 hours after the first. Barracuda rounds it to 44.

Detection came on November 4, when an engineer noticed odd activity on their own account and reported it. The same day, the vendor contacted Barracuda about its own incident. Barracuda removed the container that evening, Pacific time, and rotated credentials across every system involved. A November 13 update says two internal automation servers with exposed credentials were retired, with no sign the attacker used them.


The part customers care about: the system that stores Barracuda XDR integration credentials, which Barracuda uses to pull data from customers' Microsoft 365, Google Workspace and endpoint tools, was accessed. Barracuda found no direct evidence those credentials were viewed, but says its logging was not detailed enough to rule it out. That is why notified customers are being told to rotate.



Why Barracuda says ARTEX


The resources the attacker created had names made of the letter i plus digits, with no words or hyphens: i31911, i28324probe, i19engx1. The CI workflow files were named i19ghci.yml and i19ci2.yml, on branches starting i19ghci- and i19ci2-. The ordering of the attack and its built-in credential and wallet hunting also fit. Barracuda's advisory calls this "consistent with" ARTEX; the CISO's post says it all "match[es]" ARTEX. The CISO also makes the point that matters most: so many groups run ARTEX that the match tells you how the attacker worked, not who it was.


Barracuda reads the motive as financial (crypto and reusable credentials), and notes a third possibility it cannot rule out: since the vendor is an AI company, the target may have been AI-service credentials and the compute behind them. That is the same pattern as the Railway incident, which Barracuda calls "eerily similar" and which we covered on November 3 and November 11. Barracuda does not say Railway was its vendor, and nothing here says the two are linked.



What to hunt for


Barracuda's published indicators. It warns that seeing one does not by itself confirm malicious activity.


IP addresses: 180.178.56.10 and 52.128.238.126 (both AS963, N963 PTE. LTD.), 221.146.68.107 (Korea Telecom) and 2a09:bac1:7680:3d0::26b:5b. That last one is Cloudflare WARP, a consumer VPN shared by a very large number of people, so look for it in your logs alongside other signals but do not block it.


User agents: Boto3/1.43.122, curl/8.5.0, curl/7.81.0, Python-urllib/3.10, and Mozilla/5.0 (compatible; read-only-audit). The first four are ordinary tools and only mean something in context, such as from an access key that should never be used by a script. The last one is unusual and worth an alert on its own.


Names: cloud or Kubernetes resources named i followed only by digits and a short suffix, GitHub workflow files i19ghci.yml or i19ci2.yml, and branches starting i19ghci- or i19ci2-.


In GitHub, look for short-lived branches created and deleted within minutes that triggered a workflow, and for job logs containing long base64 or hex strings where secrets would normally be masked.



What to do


Inventory third-party keys. Every access key, service principal, token or deploy key you gave to a vendor, contractor or consultant: who has it, what it can reach, and when the project ended. Expire what is finished. Barracuda's CISO says short-lived, narrowly scoped credentials "would have stopped this intrusion at its first step."


Assume vendor repositories leak. Anything you hand a vendor can end up in their config files and their Git history. Give them credentials that expire on their own.


Cut the chain. The attacker never needed a vulnerability: each secret it read opened the next system. Stored secrets readable by the first key, CI tokens that can print secrets, Kubernetes service accounts that can schedule a container with a host mount: break any one link and the run stops. Block hostPath mounts with admission policy, and make sure a workflow on a throwaway branch cannot reach production secrets.


Log data access, not just control-plane changes. Barracuda cannot say whether its customers' integration credentials were viewed because those reads were not logged. If you could not answer the same question about your own secret store, fix that first.


Detection was a person. An engineer noticed odd activity on their own account. At machine speed, about ten accounts fell in under two days. Plan to rotate everything quickly, and practice it.



How we know, and what we did


Everything above is from Barracuda's Trust Center advisory (published November 10, updated November 13, 2026), its CISO's post "Forty-four hours" (November 11, 2026) and BleepingComputer's report of November 16. We did not see this intrusion ourselves. None of Barracuda's IP addresses were in our corpus before today, and our correlation graph found nothing next to them or overlapping the ARTEX addresses from the FBI's FLASH.


We added the three IPv4 addresses to the DugganUSA feed under source manual-batch-barracuda-artex-2026-11, credited to Barracuda, at confidence 70, below our own edge shield's blocking floor of 80, because Barracuda itself says their presence does not confirm malicious activity and one is a Korea Telecom address that may be a proxy. Pulling ips.csv with min_confidence=80 leaves them out. The Cloudflare WARP address is not in the feed.



Sources


Barracuda Trust Center, "Barracuda XDR integration credentials," published November 10, 2026, updated November 13, 2026: trust.barracuda.com/security/barracuda-xdr-integration-credentials. Arve Kjoelen, Barracuda, "Anatomy of a machine-speed intrusion," November 11, 2026: blog.barracuda.com/2026/11/11/anatomy-of-a-machine-speed-intrusion. BleepingComputer, "Barracuda links cloud breach to ARTEX AI-powered cyberattack," November 16, 2026.




Every indicator in this post is in the feed. Free.

1.58M+ IOCs, STIX 2.1 / TAXII, 88% novel vs ThreatFox, exploited-CVE leads ahead of CISA. No credit card — a free API key in 30 seconds, and you can audit every claim above against the live endpoints.


Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page