Twelve companies raided in ten minutes, 23 GB of AI models stolen in 18 minutes: Artifactory CVE-2026-82329
On 28 August 2026 JFrog published CVE-2026-82329: an authentication weakness in Artifactory that, under default configuration, lets an unauthenticated visitor obtain administrative privileges1. Four days later, admin tokens that nobody inside the companies had issued began to appear in corporate Artifactory servers. We can date that because one of the people minting them left a working directory open on the web: the scanners, a folder of notes per company, gigabytes of file listings, and a download script whose first comment calls its targets "crown-jewels". This post walks through that directory: how the raid works, what was taken, and what to look for in your own token registry.
In brief
- The whole attack is two requests. The first registers a made-up service with the server and gets a service token back. The second trades it for an admin token that is valid for one year.
- Targets came from the FOFA search engine: one run probed 11,343 hosts. Across its result files the directory records 51 Artifactory instances that handed over a token.
- On 4 September, twelve organisations were pulled in ten minutes: users, access tokens, repositories, permissions and system configuration from each. Six were then mapped in depth.
- From one consumer electronics manufacturer the operator downloaded 23.5 GB of model weights and training data in 18 minutes.
- The operator was not the first one in. Eight of the twelve instances already held admin accounts and tokens planted by other parties.
- Patching does not undo any of it. The minted tokens outlive the upgrade, and some of the planted ones never expire.
- Every script and summary in the directory describes the work as an authorised red-team engagement. Nothing else in the directory supports that.
Two requests to admin
Every script in the directory opens with the same dozen lines. They build a token that claims to be a new Artifactory service joining the cluster, sign it with the empty key a default installation accepts, and post it to the Access service. The operator's own notes call the flaw "unauth admin via blank join key". A vulnerable server answers 201 and returns a real service token. The second request presents that token and asks for another one:
POST /access/api/v1/registry/join Content-Type: text/plain -> 201, service token
POST /access/api/v1/tokens {"scope": "applied-permissions/admin", "audience": "*"}
-> 200, admin token
Both requests carry the bare user agent Mozilla/5.0. Nothing is uploaded, no account is created, and no password is guessed. The admin token that comes back is good for a year, and from that point the operator is simply a very busy administrator.
The token is not invisible, and the operator knew it. One of the summaries carries this warning to itself, and it is the most useful sentence in the directory for a defender:
OPSEC NOTE: Our own minted tokens [...] are visible in the target's token registry.
Finding the servers
Target lists came from FOFA, the internet search engine, through its paid API. The scanner states its plan and its budget in a comment:
body="artifactory" && (title="Artifactory" || title="JFrog") # 15,000 pts
title="JFrog" # 10,000 pts
body="artifactory" && port="8082" # 5,000 pts
The run log shows 11,343 hosts loaded from that pull and probed with 30 threads and an 8 second timeout. A separate list of 200 hosts on port 8082, the port Artifactory's router listens on when no proxy sits in front of it, did far better: 29 of the 200 answered with a token.
Counting only responses that contain a real token, the result files hold 39 instances, and the per-company notes add 12 more that appear in none of the scan lists. That is 51 Artifactory servers that gave an admin-capable token to this one directory.
Twelve organisations in ten minutes
On 4 September 2026, between 16:57:34 and 17:07:00 UTC, one script was run against twelve hostnames. For each it minted a token and saved the answers to eight API calls: version, full system configuration, repositories, two user lists, access tokens, permission targets and builds. The twelve are a consumer electronics manufacturer, a global IT services provider, a consumer research firm, an energy utility, an engine manufacturer, a software testing vendor, a media security vendor, a precision agriculture company, a network equipment maker, a court software supplier and two small technology firms.
They ran Artifactory versions from 7.117.8 to 7.161.17. Every one was below the fixed release for its branch1.
Six were then given what the files call a deep recon: every user, every token's metadata, every permission target, the Docker catalogue of every registry, and a recursive listing of the largest repositories.
| Organisation | Users | Access tokens | Repositories | Stored artifacts | What the directory holds |
|---|---|---|---|---|---|
| Global IT services provider | 8,986 | 5,850 | 1,091 | 553 TB | Every user record and all 1,014 permission targets, 3.1 GB of file listings, and four downloaded files: two Java keystores and two authentication libraries. |
| Consumer electronics manufacturer | 4,273 | 7,850 | 325 | 49 TB | Listings of 16 repositories (942 MB), the image catalogue of 81 container registries, and the 23.5 GB download described below. |
| Consumer research firm | 3,953 | 5,219 | 467 | 30 TB | Listings of 14 repositories, eight internal container images pulled and unpacked, and a file of credentials found inside them. |
| Energy utility (test environment) | 1,394 | 1,540 | 3,207 | 44 TB | Groups, permissions, 7,203 build records, and mail relay credentials read from the system configuration. |
| Engine manufacturer (development instance) | 137 | 210 | 214 | 244 GB | Listings of its largest repositories and the internal addresses of both cluster nodes. Nothing downloaded. |
| Software testing vendor | 653 tokens | 1 | 20 GB | Configuration and token list only. | |
A listing is not a download, but it is a floor plan: every artifact in a repository by name, size and date. On the largest instance, 181 of the 5,850 tokens carried admin scope, and the token list shows which build robots own them.
The download
For the electronics manufacturer the operator went further. The summary for that instance has a section on machine learning: a repository holding proprietary language model weights and the datasets used to train them, all "directly downloadable". A short shell script followed. With the company name removed, it begins:
# [company] crown-jewels exfil: datasets + fine-tuned model weights
...
# Priority 1: datasets (unique, irreplaceable)
# Priority 2: model weights
It starts eight curl downloads in parallel with the admin token. The log it left is a directory listing: five training datasets of 0.4 to 5.9 GB, an 11.05 GB weights file, its tokenizer and its configuration. The first file landed at 20:22 and the last at 20:40 on the server's clock, which is two hours ahead of UTC. 23.5 GB in 18 minutes, the same evening as the first pull.
Secrets inside the artifacts
The research firm shows why an artifact store is worth more than its access controls suggest. The operator pulled eight container images belonging to an internal automation team, unpacked the layers and read the files. The result is a short Markdown file of credentials: a database password and an identity provider client secret from an environment file baked into one image, a production database login written into a scheduled job definition in another, and the passphrase of an internal secrets store.
The energy utility's instance gave up mail relay credentials straight from its system configuration, which any admin token can read. From the IT services provider the operator fetched two Java keystores and two libraries whose names contain "credentials" and "token".
Not the first one in
The operator's notes on the engine manufacturer contain a finding headed "Multiple prior operators on the box". The token lists and user lists bear it out, and they put dates on it. All times are UTC.
| When | Where | What appears in the registry |
|---|---|---|
| 28 Aug 18:27 | CVE-2026-82329 is published. | |
| 1 Sep 12:05 | Engine manufacturer | The first admin token with issuer jfrt@exploit- followed by a Unix timestamp. |
| 1 Sep 22:52 | Consumer research firm | The first of 54 admin tokens with issuer jfrt@cve82329-poc-. The operator's notes mark this issuer pattern as its own. |
| 2 Sep 04:47 | Engine manufacturer, then the IT services provider at 05:37 | Tokens with subject codeb0ss_jfrt_ and ten random letters, scope admin, no expiry. 17 on one instance and 27 on the other by 4 September. |
| 2 Sep 09:08 | IT services provider, engine manufacturer | Tokens with subject yeezy_jfrt_, same shape. |
| 2 Sep 09:53 | Consumer electronics manufacturer | The operator's first token there, with two more that day. |
| 4 Sep 16:57 | All twelve | One token each with subject ending /users/jfrt@service-x: the batch pull. |
Then there are the accounts. The IT services provider's user list contains 25 accounts named nxploited_ plus three digits and one named 0xterror. All 26 are administrators in the internal realm, and none has ever logged in. Their email addresses are at pentest.local and poc.local. The same names sit on the engine manufacturer (15), the research firm (10) and the energy utility (4), where the operator's own notes list 20 planted administrators that have never logged in. Counting the first-pull files as well, eight of the twelve instances hold traces of someone other than this operator.
So within a week of the CVE record, a single exposed Artifactory server could have four or five unrelated parties holding admin on it, each leaving its own naming habit behind.
"Authorized", says every file
The deep recon script begins with the words "AUTHORIZED RED-TEAM DEEP RECON". The summaries say "authorized engagement" and describe the token as "engagement-issued". The scripts and summaries have the uniform shape of work produced by an AI coding assistant following a brief: numbered hypotheses, evidence folders, a findings list after each run.
The rest of the directory describes something else. The targets were chosen by querying a search engine for every Artifactory server it knows. The download script is titled as an exfiltration. A backup script copies the result folders to a second server. And beside the Artifactory work sits a second campaign with the same structure: a sweep of the whole IPv4 space for the WHM port, 2,990,674 addresses, tested with a public exploit for the cPanel authentication bypass CVE-2026-419403. The operator's summary counts 1,372 vulnerable servers and 34 "gold" ones holding 2,203 hosting accounts between them. A red team with a signed scope does not scan three million strangers.
The server itself is a virtualisation host at a Swiss provider with a small block of addresses, driven over SSH from a Mac. The directory was exposed by a one-line Python web server started in the home folder on 31 October.
What to check on your Artifactory
- Upgrade. The CVE record lists the fixed releases: 7.111.21, 7.117.28, 7.125.20, 7.133.29, 7.146.38 and 7.161.2012.
- Then read your token registry, because the upgrade does not empty it. List tokens as an admin (
GET /access/api/v1/tokens) and look for a subject ending/users/jfrt@service-x, an issuer startingjfrt@cve82329-orjfrt@exploit-, and subjects startingcodeb0ss_jfrt_oryeezy_jfrt_. More generally: any admin-scoped token whose issuer is not one of your own service IDs. Revoke them. - Read your user list. Look for
nxploited_and0xterror, for administrators in the internal realm who have never logged in, and for email addresses atpentest.localorpoc.local. On two instances we also saw never-used admin accounts styled as an employee address with two digits after the surname. - Search the request logs for
POST /access/api/v1/registry/joinfrom any address that is not one of your own nodes, followed byPOST /access/api/v1/tokens. Afterwards the pattern is bulk reads:/artifactory/api/security/users,/artifactory/api/system/configuration, Docker_catalogcalls on every registry, and/artifactory/api/storage/<repo>?list&deep=1. - Block the join path at the edge. Nothing outside your cluster needs to reach
/access/api/v1/registry/join. - If you find a foreign token, treat the contents as read. Rotate what the system configuration holds (mail, LDAP, remote repository credentials), then the secrets baked into images and packages. The quickest wins in this directory came from environment files and job definitions left inside container layers.
- Check integrity as well. An admin token can write. Compare checksums of release artifacts against your build system for the period the instance was exposed.
Indicators
| Indicator | Role | Confidence |
|---|---|---|
46.19.142[.]226 | The operator's server (AS51852, Switzerland): scanners, token minting, downloads. Open directory on port 8099. | High |
46.19.142[.]227 to 46.19.142[.]230 | Virtual machines on the same host, used by the operator's scripts as extra source addresses. | Medium |
179.43.175[.]98 | Second server that received copies of the result folders over rsync. | High |
Token subject ending /users/jfrt@service-x, issuer jfrt@service-x, scope applied-permissions/admin | Left by this operator's join-and-mint step on every instance it pulled. | High |
Token issuer jfrt@cve82329-poc- or jfrt@cve82329-recon- plus 12 hex characters | Marked as its own in the operator's notes. Admin scope, one year. | High |
Token issuer jfrt@exploit- plus a Unix timestamp | Another party. Seen from 1 Sep 2026 12:05 UTC. | Pattern |
Token subject codeb0ss_jfrt_ or yeezy_jfrt_ plus ten letters, scope admin, no expiry | Other parties. Seen from 2 Sep 2026. | Pattern |
Users nxploited_ plus three digits, 0xterror; admin, internal realm, never logged in; email at pentest.local or poc.local | Planted admin accounts, another party. On seven of the twelve instances. | Pattern |
POST /access/api/v1/registry/join with a text/plain body and user agent Mozilla/5.0 | The probe. A 201 response means the instance is vulnerable and now holds a foreign service token. | Pattern |
References
- CVE Program, CVE-2026-82329: potential authentication bypass leading to administrative access in Artifactory (published 28 August 2026, CVSS 9.8).
- JFrog, JFrog security advisories and Artifactory self-managed releases.
- CVE Program, CVE-2026-41940: cPanel and WHM authentication bypass via the login flow (published 29 April 2026).