-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA256
NEFARIOUSPLAN-CANONICAL-V1
{"body_md":"## The first advisory named one version. The second named every other version.\n\nOn 26 September 2026, David M. Johnson, who commits as snoopdave and wrote Roller 4.0 in 2007, posted to announce@apache.org. Subject: \"CVE-2026-82384: Apache Roller: Deserialization issue in the XML-RPC endpoint.\" The advisory body:\n\n> Affected versions:\n> - Apache Roller 6.1.5\n\nOne version. Mitigation: upgrade to 6.1.6. Finder: n0mi1k. Reporter: Matt Raible. The record went to NVD, which published it on 28 September with a `versions` array containing exactly one entry for 6.1.5.\n\nOn 17 November 2026, Arnout Engelen reissued the advisory on announce@apache.org. Subject: \"CVE-2026-129448: Apache Roller: Deserialization issue (releases before 6.1.5).\" The body:\n\n> Affected versions:\n> - Apache Roller 4.0.0 before 6.1.5\n\nFour thousand versions, speaking loosely. One additional line that the September advisory did not have:\n\n> This is the same vulnerability as CVE-2026-82384, which listed only Apache Roller 6.1.5 as affected. This issue is being actively exploited.\n\nThe reissue is doing two things at once. It is a scope correction by reissue, fixing a records field that the September CVE understated by nineteen years. And it is a disclosure after exploitation, upgrading the advisory to acknowledge in-the-wild activity that the September version did not mention.\n\nCISA received only the corrected record. On 20 November, CVE-2026-129448 appeared in the Known Exploited Vulnerabilities catalog with a due date of 23 November. CVE-2026-82384, the record that names 6.1.5, is not in the catalog.\n\n## The CPE match drops 6.1.5 by strict inequality\n\nNVD's 17 November entry for CVE-2026-129448 translates \"4.0.0 before 6.1.5\" into a CPE match rule:\n\n```json\n{\n \"vulnerable\": true,\n \"criteria\": \"cpe:2.3:a:apache:roller:*:*:*:*:*:*:*:*\",\n \"versionStartIncluding\": \"4.0\",\n \"versionEndExcluding\": \"6.1.5\"\n}\n```\n\n`versionEndExcluding` is the strict upper bound. The CPE grammar treats this as a half-open interval: `[4.0, 6.1.5)`. Version 6.1.5 is on the open side. An automated scanner matching installed versions against this rule will not flag a Roller 6.1.5 host.\n\n6.1.5 is covered by the September CVE, which exists, which NVD still serves, and which CISA did not add to KEV. The two records together describe the full affected range. Neither one describes it alone.\n\n## 6.1.5 is the version the researcher reported against\n\nn0mi1k reported the bug against Roller 6.1.5. The fix shipped in 6.1.6. The September record, written in the usual CNA shorthand of \"affected version is the version that got patched,\" listed 6.1.5 and stopped. That reading would be correct for a regression introduced in 6.1.5 and fixed in 6.1.6. It is not correct here.\n\nThe misconfigured servlet is the XML-RPC endpoint. `WEB-INF/web.xml`:\n\n```xml\n\n XmlRpcServlet\n org.apache.xmlrpc.webserver.XmlRpcServlet\n \n enabledForExtensions\n true\n \n\n```\n\nThat init-param turns on Apache ws-xmlrpc's vendor extension parsing, including the `ex:serializable` type. Git blame on `apps/weblogger/web/WEB-INF/web.xml`: the line was written on 13 August 2007 by snoopdave, in the commit that landed Roller 4.0 RC1. The file is identical across the tags for roller-4.0, roller-5.0, roller-5.1.2, roller-5.2.0-final, roller-6.0.0, and roller-6.1.5. Nineteen years, six major versions, same line, no changes.\n\nThe CPE rule's lower bound, 4.0, is correct to the day. The exclusion of 6.1.5 from CVE-2026-129448 is not a code fact. It is an artefact of CNA bookkeeping: the CVE record is immutable by design, so the two CVE records cannot overlap in their affected ranges because that would double-count. CVE-2026-129448 takes everything up to but not including the version the first CVE claimed. Scope correction by reissue is the protocol's only move when the first number was right about a bug and wrong about its reach.\n\n## The parser does not read the version number\n\nThe sink lives in Apache ws-xmlrpc 3.1.3, the library Roller bundles. `org.apache.xmlrpc.parser.SerializableParser.getResult`:\n\n```java\nByteArrayInputStream bais = new ByteArrayInputStream(bytes);\nObjectInputStream ois = new ObjectInputStream(bais);\nreturn ois.readObject();\n```\n\nThe SAX parser reaches this call during request parsing, before Roller's authentication filter, before handler dispatch, before the `webservices.enableXmlRpc` admin toggle is consulted. The mechanism was walked through on this site last month in [the post on CVE-2026-82384](/posts/apache-roller-cve-2026-82384-toggle-was-never-pre-auth): the toggle gates method invocation, not parsing, and the parser runs whenever the servlet is mapped. The code path does not know what Roller version is running on top of it.\n\nThis is the receiver-supports-no-sender shape. Roller's own client code and its test suite never emit an `ex:serializable` element on the wire. The Blogger API and MetaWeblog API clients Roller ships use only standard XML-RPC types. The decoder for the vendor extension sits in the library because ws-xmlrpc shipped with it in 2007, and Roller enabled it in `web.xml` the same year, and nothing in Roller has ever asked the library to decode one. The code path is exclusively exercised by attacker traffic that sets the trigger field.\n\nWhat the reissue clarifies is the scope of exposure. \"Every production Roller deploy since 2007\" is a different asset-management problem than \"whoever upgraded to 6.1.5 in the three months between its release and the September advisory.\" The first is the problem the attackers actually have a target list for. The second is the problem the September advisory described.\n\n## Three independent PoCs, three days, the same probe\n\nn0mi1k released a scanner on the day of the September disclosure. Two more independent implementations followed the 17 November reissue.\n\nYanis Alim of TUXSHIN SECURITY published `Yanixos/APACHE-CVE-2026-129448` on 18 November: a Python probe plus a `CommonsBeanutils1` exploit that generates its gadget via a Dockerized ysoserial against JDK 11. The probe oracle sends a serialized `lab.ProbeGhost`, a class deliberately absent from any real server, and classifies the response:\n\n```\nVULNERABLE -> \"Failed to load class for result object: lab.ProbeGhost\"\nPATCHED -> \"The tag {...}serializable is invalid, if isEnabledForExtensions() == false.\"\n```\n\nA second author published `evenomn/CVE-2026-129448` on 20 November. Different gadget assembly, in-house javassist generator, no ysoserial dependency. The repository notes a wrinkle the first author's write-up does not: the `xmlns:ex` declaration has to live on the `` element. Several earlier public drafts had placed it inline on ``, and ws-xmlrpc 3.1.3 rejects that form with \"The prefix ex isn't the prefix, which has been defined last.\" The two independent scanners agree on probe shape, on expected vulnerable response, and on which Roller versions respond that way. Both verified against the official 5.2.0 WAR. Both verified against 6.1.4. Neither write-up treats 6.1.5 as a special case.\n\nThe NVD record for 6.1.4 and the NVD record for 6.1.5 are separate documents. The code path they describe is the same code path.\n\n## The catalog is a chronology. The bug is one config line.\n\nThe CVE record is immutable by design. Once a number is assigned and the record is published, the affected-versions field is a statement of record; later refinements come as new records, not edits. This is the correct protocol for provenance and poor protocol for communicating scope changes. An analyst following the September advisory saw one version. An analyst following the November advisory sees every other version. The two assertions are not contradictory in CVE-record semantics. They are both true. They are describing a nineteen-year-old `web.xml` line that neither record references.\n\nThe disclosure after exploitation note attached to the November advisory is what moved the record into the KEV catalog. The September advisory's assertion of 6.1.5 as the affected version did not get the same treatment, because KEV follows CVE-ID-level evidence, not bug-level evidence. For an admin reading the KEV list on 21 November, the only Roller entry is the one that excludes 6.1.5 by strict inequality.\n\nPoC: [Yanixos/APACHE-CVE-2026-129448](https://github.com/Yanixos/APACHE-CVE-2026-129448), [evenomn/CVE-2026-129448](https://github.com/evenomn/CVE-2026-129448).","closing_line":"The first CVE lists 6.1.5. The second CVE lists every version except 6.1.5. KEV lists only the second.","hook_md":"Apache's 17 November advisory names the vulnerability. The subject line is \"CVE-2026-129448: Apache Roller: Deserialization issue (releases before 6.1.5).\" The affected-versions line reads \"Apache Roller 4.0.0 before 6.1.5.\" The NVD record published the same day resolves that phrase into a CPE match with `versionEndExcluding` set to 6.1.5, a strict inequality that excludes 6.1.5 from the advertised range. A second CVE, issued in September, named only 6.1.5. CISA added the November one to the Known Exploited Vulnerabilities catalog on 20 November and did not add the September one. An administrator running Roller 6.1.5 who opens the CVE CISA listed sees a version range they are not in.","post_id":887,"slug":"apache-roller-cve-2026-129448-second-cve-excludes-6-1-5","title":"CVE-2026-129448: Apache's Second Roller CVE Excludes 6.1.5 By Name. CISA's KEV Catalog Lists Only The Second CVE.","type":"initial","unreadable_sentence":"An administrator running Roller 6.1.5 who consults the only CVE CISA listed for this bug sees a version range they are not in."}
-----BEGIN PGP SIGNATURE-----
iHUEARYIAB0WIQRf0htP5+SjynlxywneZjl4jgkQJgUCawHyJQAKCRDeZjl4jgkQ
JjtiAP4tYTVXfhYQRF4iBMgNlhO5QBzmK/KYeeVURDj5mC1jWgD+KklqQmnmaVhs
f5UPI4g9/IcZwPVN69r3iZFbr6wwQgs=
=NVfy
-----END PGP SIGNATURE-----