Java vulnerabilities of the week: 12 to 18 November 2026

When a vulnerability goes public you get an identifier and a number, and neither one tells you what to change. The identifier lets you track it and the number tells you how loudly to argue for a maintenance window, but the thing you actually learn from, the part that moves from someone else’s codebase into the way you write your own, is the mechanism. That is what I try to pull out each week. The longer form of the argument is in the CVE stigma post.

This is 12 to 18 November 2026, and it is an authentication week. Apache Shiro published eight advisories, Apache TomEE consolidated thirty-eight, and Apache Roller, Apache WSS4J, Apache Griffin and Apache SeaTunnel Web each shipped an auth or authorization issue. Two shapes recur: the check that defaults to success when it is skipped, and the work that happens before the request is authenticated at all. The headline sits slightly apart from both, a memory-corruption bug in OpenJDK’s Base64 decoder that is remote code execution without credentials, and it belongs to the week anyway because Base64 decoding runs in front of authentication. The deep dive I take here is Shiro, where more than one realm would say yes to a password it never checked.

The week in five

Two honourable mentions, both in retired projects. Apache Griffin’s web service has no authentication on its REST API at all, including job submission to a Livy/Spark backend (CVE-2026-128841, all versions, no fix planned). And Apache SeaTunnel Web 1.0.2 lets a disabled account log in with its valid credentials and collect a fresh token, because the login query never checks account status (CVE-2026-128981, the account-status check only reached the main branch, with no release carrying it). Neither will get a release, so if either is in your stack, restrict access or replace it.

OpenJDK Base64: where Java’s memory safety ends

Java’s memory safety is a property of bytecode over the managed heap, and the JIT steps outside it. HotSpot’s C2 compiler replaces hot methods with intrinsics, and the fastest of those are hand-written assembly; Base64.Decoder.decode is one, built on AVX-512 VBMI instructions on x86-64 CPUs that support them. The intrinsic computes a 64-bit write mask from the input and output lengths and then adjusts it for = padding, and for a tail that decodes to nothing the adjustment underflows the length and turns the mask to all ones. So a 65-byte input of 64 A characters and one = sizes a 50-byte output buffer and then writes 64 bytes into it, 62 of them past the end, on top of the next object on the heap. The decode call throws afterwards, but the neighbour is already overwritten. From that fixed overwrite the finder grooms the heap into a chosen-value write and a Heartbleed-style leak, and on to remote code execution, which because Base64 decoding runs before authentication (Basic headers, JWT segments, SAML) is pre-auth in Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate.

The lesson is the boundary: “Java is memory-safe” describes your bytecode, not the native intrinsic the JIT swaps in underneath it, and you will not find this by reading your own Base64.getDecoder().decode(s). The fix is a runtime upgrade, the November 2026 CPU, not a dependency bump; if you cannot apply it at once, -XX:-UseBASE64Intrinsics or -XX:UseAVX=2 on x86-64 takes you off the vulnerable path. It affects OpenJDK 17.0.6 and later on AVX-512 VBMI CPUs; 8 and 11 are not affected.

Apache Shiro: authentication has to fail closed

Shiro is a security framework for Java applications: it authenticates a subject, authorizes what that subject can do, manages the session and handles the crypto underneath. This week it published eight advisories, fixed across 3.0.1 and 3.0.2, and the two that carry the lesson are both about a credential check that produces “authenticated” without having verified anything.

The serious one is CVE-2026-129349, CVSS 4.0 9.1. DefaultLdapRealm (and JndiLdapRealm) authenticate by binding to the LDAP server with the submitted username and password; a successful bind is the proof. The realm can also cache a successful authentication, which is off by default but a common performance switch. With caching on, the first login does the bind and caches the result, and the second login for that user is served from the cache, which skips the bind. What is left to check the password is the CredentialsMatcher, and the one the realm installs by default is AllowAllCredentialsMatcher, which returns a match for any credentials at all. So after a user logs in once, anyone who knows that username authenticates as them with any password. The fix in 3.0.1 stops installing AllowAllCredentialsMatcher by default. On older lines the advisory’s mitigations are to disable authentication caching on the LDAP realm (the default anyway) or to configure a real matcher such as SimpleCredentialsMatcher.

The quiet one is CVE-2026-129347, CVSS 4.0 2.3, and its the same idea with a smaller blast radius. DefaultPasswordService and SimpleCredentialsMatcher could treat an empty stored credential as a match for an empty submitted password, so any account holding an empty credential can be logged into with an empty password. It only bites deployments that have such accounts, which is why the score is low, but the mechanism is the one above: nothing is verified, and the outcome is a match rather than a rejection. It was found by the ASF using an LLM security scan, and the fix is in 3.0.2.

The other six fill in the surface. CVE-2026-74762 is session fixation through RememberMe: a subject restored from a remember-me cookie kept its existing session rather than getting a fresh ID (fixed in 3.0.1). CVE-2026-74763 is authorization annotations such as @RequiresRoles not always enforced depending on whether they sit on an interface, a class or a superclass, across the shiro-guice, shiro-aspectj, shiro-cdi and shiro-jakarta-ee modules (fixed in 3.0.2, the annotation nearest the invoked method now winning). The rest are a PropertiesRealm reload that keeps removed users and widened permissions (CVE-2026-129348), a saved-form replay that can weaken a resubmitted form’s cookie attributes (CVE-2026-88019), the shiro:principal tag not escaping its output (CVE-2026-129320, a stored XSS found by the ASF using Claude agents), and an incomplete fix for an earlier path-matching default (CVE-2026-114737).

Worth noticing that the same week lands the same class in a different stack. TomEE’s CVE-2026-111261 is an authentication bypass with an empty password in its LDAP identity store, and CVE-2026-111240 is authentication for accounts with no stored password in its SQL login module. Same class, two Java stacks, one week.

Why I find it interesting

This is CWE-287, improper authentication, and the specific family is fail-open: when the step that verifies a secret is skipped or short-circuited, the result defaults to success instead of failure. The LDAP case skips the bind through the cache and then accepts anything through the default matcher. The empty-credential case compares nothing to nothing and calls it a match. Both arrive at “authenticated” along a path where no secret was actually checked.

The place this hides is caches and defaults. A cache exists to skip work, and if the work it skips is the credential check, the cached “authenticated” has to carry the fact that the credential was already verified, or the skip becomes a bypass. The review question is concrete and worth running against your own login code: for every path that can return “authenticated,” what is the result when the submitted secret is empty, or when the verifying step did not run? If it is anything other than rejection, that is the finding.

One practical note on reach, stated plainly because it decides who has work to do. Most of these affect all Shiro versions through 3.0.0 or 3.0.1, and that phrasing on the project’s security page takes in the 1.x and 2.x lines, not only 3.0. The fixes are in 3.0.1 and 3.0.2, and on Maven Central the newest 1.x release is 1.13.0 and the newest 2.x is 2.2.1, so for those lines clearing the core issues means moving to 3.0.2. In the meantime the per-CVE configuration mitigations (caching off on the LDAP realm, a real matcher, annotations on the implementing method, a restart after editing a properties realm) are what you have.

What I take away from this week

First, two auth patterns did all the work. One is fail-open, which is the Shiro and TomEE thread above: the empty or absent credential accepted, the bind skipped, the annotation not enforced, the EJB method authorization skipped on JDK 21 and later. The other is work before authentication, which is Roller deserializing an XML-RPC request before it checks the caller or the feature flag, and WSS4J validating a message before it records that it has seen it. Both come down to ordering. The verification has to happen before the effect, and its result has to default to no. When either of those slips, the code still runs; it just runs for the wrong person.

Second, where the fixes actually reach people, because they do not all reach the same way. OpenJDK reaches everyone, because it is the runtime and not a library: no BOM bump protects you, the JDK sits below the BOM, so this is a platform upgrade (the November CPU) or a startup flag. WSS4J travels through Apache CXF, which carries it and turns the replay cache on by default, so a CXF bump to a release pulling 2.4.5, 3.0.7 or 4.0.3 is how most people get it. TomEE is a direct upgrade to 10.3.0 or 11.0.0-M2, with twelve of the thirty-eight in removed modules you turn off rather than patch. Shiro is 3.0.2, with the version-line caveat above. Roller is an application upgrade to 6.1.6 and the one to do now, because it is being exploited in the open. Griffin and SeaTunnel Web are retired: restrict access or replace.

Third, briefly, because I noted it last week too: AI turned up on both sides again. The OpenJDK bug was found and weaponised with AI assistance by its finder, who puts the discovery at about an hour, and two of Shiro’s advisories credit an LLM scan and Claude agents on the finding side. Last week it was the Sling and Axis advisories. I will keep leaving it as the fact it is, rather then drawing a conclusion from two weeks.

References