Status, November 17, 2026, 22:15 UTC. CVE-2026-114938 is reserved but empty: no description, no score, no affected versions, and nothing in NVD. Oracle, OpenJDK and the major Java distributions have not confirmed it or tied it to a fix. No public exploit and no attacks have been reported. We will update this post as that changes.
- The claim: a pre-auth heap overflow in the AVX-512 Base64 decoder in HotSpot, the JVM inside OpenJDK 17.0.6 and later (some reports add 21 and 25). A crafted Base64 string in an auth header, token or upload would corrupt memory before any login check.
- What is confirmed: the CVE number is reserved. That is all. The root cause, the version range, the score and the fix are all secondhand and the sources disagree.
- What the source code says: the AVX-512 decoder only runs on CPUs with AVX-512 VBMI (Intel Ice Lake and newer, AMD Zen 4 and newer). Arm servers, Apple silicon and older Xeons (Skylake, Cascade Lake) never take that path. On Java 17 it only handles input over 128 bytes; on Java 21 and 25, input from 32 bytes.
- Do now: inventory Java services that decode untrusted Base64 before authentication, check their CPUs, and be ready to patch. The interim flag
-XX:-UseBASE64Intrinsicsturns the vector decoder off, at a speed cost. - Watch for: JVM crash logs (
hs_err_pid*.log), Java containers exiting with 134 or 139, and Base64 tokens whose length is one more than a multiple of four. Details below.
What Is Being Claimed
On November 17 the name "Java4Shell" spread fast, with CVE-2026-114938 attached. The story, as relayed by Mallory and others from Oligo's research: an integer underflow in the last step of HotSpot's AVX-512 Base64 decoder lets a short, crafted input (65 bytes in the report) write 62 bytes past the end of a heap buffer. Datadog describes it as an out-of-bounds write that happens before the application can reject the input. South Korea's NCSC notice, as reported in the Korean press, puts the range at OpenJDK 17.0.6 and later and gives it CVSS 4.0 9.2.
Base64 is everywhere in front of a login: HTTP Basic auth, JWTs, SAML responses, API keys, file uploads. If the bug is real and reachable, an attacker would not need an account. That is why people are reaching for the Log4Shell comparison.
What Is Confirmed and What Is Not
| Claim | Status |
|---|---|
| CVE-2026-114938 exists | Reserved only. No description, score or affected versions. NVD, Red Hat, Debian and Ubuntu have nothing. |
| Root cause is the AVX-512 Base64 decoder | Relayed by an aggregator and credited to Oligo, but DugganUSA could not find it in Oligo's own post, and Datadog describes it differently. Not confirmed by Oracle or OpenJDK. |
| OpenJDK 17.0.6 and later is affected | From a relayed government notice. The AVX-512 decoder change (JDK-8268276) shows up in the 17.0.7 release notes, not 17.0.6. |
| Java 21 and 25 are affected | They contain the same decoder. No source has confirmed them. |
| Oracle's November 17 Java update fixes it | Oracle shipped 11 Java SE fixes (highest CVSS 3.1 score 8.1) without naming CVEs, so nobody can say yet whether this is one of them. DugganUSA traced every "9.2 critical" report back to a single origin. |
| Exploited in the wild | No reports. Not in CISA KEV. |
| Public proof of concept | None published. |
What the OpenJDK Source Tells Us
We read the current update sources for Java 17, 21 and 25 (stubGenerator_x86_64.cpp in openjdk/jdk17u, jdk21u and jdk25u). Two conditions narrow the exposure, if the reported root cause is right:
- The vector decoder needs AVX-512 VBMI and BW. The code only builds the AVX-512 path when the CPU reports both. Plain AVX-512 is not enough. That rules out Skylake and Cascade Lake Xeons, which have AVX-512 but not VBMI, along with every Arm server and Apple silicon. It leaves Intel Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD Zen 4 and Zen 5 (EPYC Genoa, Bergamo and Turin). Much of today's cloud fleet is on those.
- The input has to be long enough. On Java 17 the decoder sends anything of 128 bytes or less to the plain scalar loop, so a bare 65-byte string never reaches the AVX-512 code there: either the reported 65 bytes is what remains after the vector loop on a longer input, or the Java 17 claim is off. On Java 21 and 25 the cutoff is 32 bytes, so a 65-byte input does take the AVX-512 path. That makes 21 and 25 at least as worth checking as 17, even though the public reports lead with 17. (21 and 25 also have an AVX2 decoder; nobody has implicated it.)
We have not reproduced the bug. This only tells you which hosts could even run the code in question.
How to Check Your Hosts
- Does the CPU have VBMI? On Linux:
grep -o -w avx512vbmi /proc/cpuinfo | head -1. No output means the vector decoder cannot run on that host. Check the CPU of each node in a cloud or Kubernetes pool; instance families differ. - Which Java does each service run?
jcmd -llists running JVMs, andjcmd <pid> VM.versionprints the build. Do not forget runtimes bundled inside products (gateways, SIEM and HMI platforms, IDE servers) and inside container images. - Is the intrinsic on?
java -XX:+PrintFlagsFinal -version | grep UseBASE64Intrinsics. It is on by default. - Does the service decode untrusted Base64 before auth? Basic auth, bearer tokens, SAML, OAuth callbacks and upload endpoints facing the internet come first.
Patching: Which Builds Exist
Oracle's November 17 builds are out: 8u513 (1.8.0_513-b04), 11.0.33.1+4, 17.0.21.1+4, 21.0.13.1+4, 25.0.5.1+4 and 27.0.1.1. Per DugganUSA, as of 19:31 UTC Eclipse Temurin and Amazon Corretto had not shipped matching builds; their newest were 17.0.21 and 21.0.13 from October. An Oracle download does not patch a Temurin, Corretto, Zulu, Red Hat or Microsoft runtime: each runtime has to come from the vendor that built it, so find out who built yours. Oracle's builds also turn on a Kerberos cross-realm transited-field check by default, so test before rolling them out where Java does Kerberos.
Install the updates you have. Just do not assume they close Java4Shell until a vendor says so.
The Interim Workaround
Starting the JVM with -XX:-UseBASE64Intrinsics turns off the hand-written decoder, so Java falls back to the plain implementation. Set it on the command line or through JAVA_TOOL_OPTIONS; the JVM prints Picked up JAVA_TOOL_OPTIONS at startup, so you can confirm it took. It needs a restart, Base64-heavy workloads get slower, and no Java vendor has recommended it yet. It only helps if the reported root cause is right. On hosts without VBMI it changes nothing, so start with the hosts that have it.
How to Spot It if It Is Real
There are no published payloads or IOCs. These are the signals we would watch, based on how a heap overflow in the JVM behaves and on the reported trigger. We are keeping them out of our detection feed while the bug is unconfirmed, because a rule built on an unverified description is guesswork that pages people at 3 a.m.
- JVM crash logs. A HotSpot crash writes
hs_err_pid<pid>.login the working directory, or where-XX:ErrorFilepoints. Watch for new ones on internet-facing services. In the file, theProblematic frameline names where it died. A frame inside a stub routine with "base64" in its name, or heap corruption reported right after a Base64-heavy request, is the pattern to escalate. - Containers and services dying without a deploy. A JVM that aborts exits with 134 (SIGABRT); a raw segfault is 139 (SIGSEGV). In Kubernetes, look for restarts with those exit codes and no rollout. On systemd, look for
status=6/ABRTorstatus=11/SEGV. One crash on one pod is noise; a cluster of crashes across replicas behind the same endpoint is not. - The request just before the crash. Line up the crash time with the web server, load balancer or WAF log and look at the last requests to that instance: an unusually long
Authorizationheader, a malformed JWT, an odd SAML POST. - Base64 that no encoder would produce. Valid Base64, padded or not, never has a length that is one more than a multiple of four. A 65-byte value is one of those. A Basic auth credential, JWT segment or SAML blob with length mod 4 equal to 1 is malformed by construction, and worth logging even if nothing crashes. It is cheap to compute at a proxy or WAF.
- Your own mitigation.
Picked up JAVA_TOOL_OPTIONS: -XX:-UseBASE64Intrinsicsin startup logs tells you which services have the workaround. That is a configuration check, not an attack signal.
What We Did on Our Own Systems
We ran the checks above across our own Java services today. Most of them sit on CPUs without AVX-512 VBMI, so the vector decoder cannot run there. The one host with AVX-512 has no VBMI either, and its Java ports are closed to the internet. We will patch to the vendors' fixed builds once they confirm which ones those are.
RedEye Assessment
Treat it as plausible and unconfirmed. A memory bug in a hand-written vector routine is entirely believable, and the reach would be large if it holds: Base64 sits in front of authentication in almost every Java web stack. But as of tonight there is one primary research claim, a relayed government notice with a version range that does not match the source history, an empty CVE record, and no vendor statement. That is not a Log4Shell yet.
The useful work today is inventory: which Java services decode untrusted Base64 before auth, and which of them run on VBMI-capable CPUs. That list is short in most environments, and it is the list you will want the moment a fix or an exploit lands. We will add feed detections when the bug is confirmed and the trigger is known.
Sources: DugganUSA, Fenko Security, watchTowr, JFrog, Endor Labs, Equixly, Mallory, OpenJDK jdk17u source.
Questions about your exposure?
RedEye Security provides assessments for organizations that need to understand their real risk.
Talk to us