Apache Roller CVE-2026-129448: unauthenticated XML-RPC RCE, exploited
Apache Roller 4.0.0–6.1.4: unauthenticated XML-RPC deserialization RCE (CVE-2026-129448, CVSS 9.8), exploited in the wild. Patch 6.1.6; no public PoC yet.
Published by Zero Hunt, an autonomous AI red team on an on-premise appliance running private AI: automated penetration testing for networks and infrastructure, black-box or gray-box, with a human approving every step that matters.
The Apache Software Foundation says an unauthenticated remote-code-execution flaw in its Roller blog server is being exploited right now. CVE-2026-129448 (CVSS 9.8) is a deserialization bug in the XML-RPC endpoint, and the ugly part is not the score — it is that two things you would normally trust to tell you whether you are exposed both lie here. The admin-UI toggle that says XML-RPC is "disabled" does not close the hole, and the first advisory for this bug, published in September, named the wrong versions.
Developing story — first published 20:34 CET (19:34 UTC), 18 November 2026. Updated as the vendor and CISA publish more.
At a glance
| CVE | CVE-2026-129448 |
| Product / affected versions | Apache Roller 4.0.0 through 6.1.4 · XML-RPC endpoint (/roller-services/xmlrpc) |
| Fixed in | 6.1.6 (extension types disabled); 6.1.7 released 2026-11-11 rolls up 15 fixes |
| CVSS | 9.8 critical · CVSS 3.1 · AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H · scored by Apache ([email protected]); NVD not yet scored (status Received) |
| Exploited in the wild | Yes — the ASF advisory (2026-11-17) states the issue "is being actively exploited" |
| CISA KEV | Not listed (catalog version 2026.11.17) |
| Public PoC | No public PoC located as of 2026-11-18; finder n0mi1k, reporter Matt Raible |
| Official advisory | ASF advisory, oss-security 2026-11-17 |
What happened: the XML-RPC endpoint deserializes before it authenticates
Roller ships an XML-RPC API, mapped by default at /roller-services/xmlrpc, built on the Apache XML-RPC library. That library supports a set of non-standard "vendor extension types" — extra serialized value types beyond the XML-RPC standard. Roller never uses them; it only ever needs the standard value types. But the servlet was configured to accept them, and it deserialized those extension types during request parsing — before it checked who was asking. That is the whole bug: Deserialization of Untrusted Data (CWE-502) reachable by anyone who can send an HTTP POST, no credentials, no interaction.
Worse, the servlet is mapped unconditionally in web.xml. So the parse-and-deserialize step runs even when the XML-RPC feature is switched off in the administration UI — older versions deserialize the request first and consult the "is XML-RPC enabled?" setting only afterward. Turning it off in the dashboard therefore does nothing against this path.
The upstream fix (PR #171, merged 2026-09-05) does two separate things: it disables the vendor extension types outright, and it gates the mapped endpoint on the real feature flag, so a disabled API is closed at the endpoint instead of behind a per-handler check that runs too late.
Who is exposed — and why your version check may have misled you
This is the same vulnerability first assigned CVE-2026-82384 back in late September. That original record listed only Apache Roller 6.1.5 as affected. A lot of operators read it, saw "6.1.5," checked that they were on some older line, and moved on.
They were wrong. The corrected advisory re-issues the finding as CVE-2026-129448 with the real affected range: Apache Roller 4.0.0 through 6.1.4 — plus 6.1.5, which the earlier record already covered. Roller 4.0.0 dates to 2008. The same unauthenticated deserialization path has been reachable for roughly seventeen years of releases.
"I checked in September. We weren't on 6.1.5, so the advisory didn't apply to us."
It applied. The September record undercounted the affected versions. If your Roller is anywhere from 4.0.0 to 6.1.4 and the XML-RPC servlet is mapped — which is the default — you are in scope, and the version number you checked is the reason you thought you weren't.
Roller is a Java blog server that has historically run high-visibility corporate and institutional blog farms, so the exposed instances tend to be internet-facing by design. The finder is credited as n0mi1k; the ASF reporter is Matt Raible.
Remediation
1. Am I affected?
Check the running version and, more importantly, whether the endpoint answers:
# Does the XML-RPC servlet respond at all? (default mapping)
curl -s -o /dev/null -w '%{http_code}\n' \
-X POST https://your-roller-host/roller-services/xmlrpc \
-H 'Content-Type: text/xml' \
--data '<?xml version="1.0"?><methodCall><methodName>system.listMethods</methodName><params></params></methodCall>'
Any Roller in 4.0.0–6.1.4 with that servlet mapped is affected regardless of the admin-UI XML-RPC setting. Do not treat a "disabled in the dashboard" status as a mitigation — it isn't one for this path.
2. Patch — exact fixed versions
Upgrade to Apache Roller 6.1.6 or later. Prefer 6.1.7 (released 2026-11-11), which rolls up fifteen fixes including this one, and restart Roller afterward.
3. Can't patch right now? Compensating controls
Per the ASF advisory, before you can upgrade, do one of:
- Set
enabledForExtensionstofalseinWEB-INF/web.xml, or - Remove the
/roller-services/xmlrpcservlet mapping,
and restart Roller. Both close the deserialization path. Do not rely on disabling XML-RPC in the administration UI — the servlet deserializes the request before it consults that setting. As a second layer, restrict the endpoint at the reverse proxy or WAF to known management sources.
4. Hunt for compromise
No vendor IOCs have been published, so hunt on behavior, not hashes:
- POST requests to
/roller-services/xmlrpc, especially large or base64-heavy bodies andmethodCallpayloads carrying the library's extension-type markers — map to T1190 (Exploit Public-Facing Application). - The Roller JVM spawning a shell (
/bin/sh,bash,cmd.exe) or other unexpected child processes — T1059 (Command and Scripting Interpreter). - Unexpected outbound connections from the Roller host to never-seen destinations shortly after such a POST — staging, beacon, or second-stage pull.
Correlate the first signal (the request) with the second and third (execution, egress). A deserialization hit that reaches code execution is loud on the wire even when the request itself looked like ordinary API traffic.
5. Eradicate and verify
If you find evidence of execution, assume the Roller process context is compromised: rebuild or clean the host, rotate the database credentials and any secrets readable by the Roller user, and only then confirm the fix. Verification is not "the version string changed" — it is that a crafted extension-type XML-RPC request is now rejected at the endpoint, not quietly deserialized. Re-send the probe and confirm the endpoint refuses it.
What is not known yet
- Who is exploiting it, at what scale, with which indicators. The advisory confirms active exploitation but names no threat actor, victim count, or IOCs.
- NVD scoring. NVD still lists the record as Received; the 9.8 is Apache's own CNA score and NVD may add its own.
- CISA KEV. Not listed as of catalog 2026.11.17. An unauthenticated, exploited-in-the-wild RCE is a strong KEV candidate; expect it to appear.
- A public PoC. None located as of this writing, though the advisory asserts exploitation is already happening — so absence of a public PoC is not safety.
When the version string and the config both lie, you have to drive the endpoint
The trap in this CVE is epistemic, not technical. Two signals you would normally trust each tell you the wrong thing: the admin-UI toggle says XML-RPC is off while the endpoint still deserializes, and the September CVE said "6.1.5 only" while the real code path reaches back to 2008. You cannot answer "am I actually exposed?" from a version number or a config flag. You answer it by sending a request and watching what the running code does.
That is exactly where an autonomous AI red team earns its place. Zero Hunt runs gray-box against your own estate on an on-premise appliance: it fingerprints the exact installed Roller build and configuration, then sends a benign extension-type XML-RPC request to /roller-services/xmlrpc and observes whether the bytes are deserialized before authentication on your instance — a working proof per environment, generated per target by local models, executed in an ephemeral hardened container that never touches the appliance host. Because the models run in the rack, your configuration and credentials never leave to a third party, and a human approves the action that matters. The result is not "you are running 6.1.3, which the advisory lists" — it is "this endpoint deserializes your input before it checks who you are," or the opposite, verified. For the fix itself, the remediation path above is a sequencing problem a defender can win: patch to 6.1.7, drop the web.xml stopgap in for the minutes before the restart, and re-verify with the same proof that demonstrated the exposure — the Remediation Advisor ranks the work by real exploitability (exploited-in-the-wild, here) and closes the loop by re-running the probe until the endpoint rejects it.
The on-premise, human-in-the-loop model matters most on days like this: the question is not what a catalog says about a version, but what your server does when someone POSTs to it — and that is a thing you test, not a thing you look up.
Every vulnerability CISA lists as exploited, with federal due dates: CISA KEV tracker →
Patching one CVE at a time? A CTEM program validates which exposures are actually exploitable before the next advisory: what CTEM is →
Is this exploitable in your environment?
Zero Hunt answers that on your own network: an autonomous AI red team on an on-premise appliance, running on private AI, black-box or gray-box, with a human approving every step that matters. Proof of what is exploitable, the fix, and signed evidence — no data leaves your perimeter.