Resources › CVE intelligence › Java4Shell
Java4Shell
CVE-2026-114938
Verdict
Java4Shell is a heap write past the end of an array in OpenJDK's x86-64 Base64 decoder, reachable without authentication. A machine is exposed only when all four requirements below hold; patched builds are 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1 and later, but version strings mislead both ways, and GraalVM 25 releases carry their own unpatched copy.
Does it reach you?
Exposed = all four hold. If any one is false for a machine, it's out.
| Requirement | How to check it | Rules a machine out when |
|---|---|---|
| 1. The CPU has AVX-512 VBMI, plus AVX-512 BW and BMI2, as the JVM sees them | grep -o -w -E 'avx512f|avx512bw|avx512vbmi|bmi2' /proc/cpuinfo | sort -u | Any of the four is missing. Intel Ice Lake, Sapphire Rapids and newer and AMD Zen 4 and newer have them; Skylake-SP, Cascade Lake and Cooper Lake don't (we ran the unpatched JVM on one: no out-of-range write). A hypervisor that hides VBMI rules the guest out |
| 2. The Java build lacks the patch | The build itself, not the version string (table below) | The patch is in the build: 17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1 and later, or a distribution build carrying it. Standard Java 8 and 11, and 17.0.0 to 17.0.5, never had this decoder |
| 3. The running JVM leaves the fast Base64 decoder on (the default) | jcmd PID VM.flags -all: UseAVX 3 or higher and UseBASE64Intrinsics true | -XX:-UseBASE64Intrinsics, -XX:UseAVX=2 or lower, or C1 only (-XX:TieredStopAtLevel=1). Read the running process: environment variables and later options can override the command line |
4. Network input reaches java.util.Base64's basic or URL decoder | Where the application decodes Base64 from requests | Only the MIME decoder or Decoder.wrap() is used, or the libraries decode with their own code (Undertow, Netty, gRPC, jjwt, Nimbus, OpenSAML). HTTP Basic authentication in Tomcat, Jetty and Spring Security, auth0's java-jwt and the Spring Session cookie do reach it, before login |
Not required, so these don't rule a machine out: running in a container (containers see the host CPU), the operating system (Linux, Windows and macOS on x86-64 share the code), or warm-up (an attacker can warm the decoder with valid input first). Inputs from 65 bytes reach it on Java 21 and later, from 129 bytes on Java 17.
Are you actually fixed?
Fixed versions by build. Where a version number misleads, the row says so.
| Build | Status as of 19 November 2026 |
|---|---|
| Upstream OpenJDK 17.0.21.1+2, 21.0.13.1+2, 25.0.5.1+2, 27.0.1.1+4 and later | Fixed |
| Builds of the same versions at +0 or +1 (for example 21.0.13.0.3) | Not fixed. The version looks current; the patch isn't in it |
| Upstream 8, 11, and 17.0.0 to 17.0.5 | Not affected: no such decoder |
| End-of-life Java 18 to 20, 22 to 24 and 26 (including Amazon Corretto 22, 23, 24, 26: "No Fix Planned") | Not fixed, and won't be |
| Debian 12 bookworm (DLA-4917-1), Debian 13 trixie (DSA-6639-1, DSA-6640-1), sid openjdk-27 | Fixed |
| Debian forky openjdk-21 21.0.13.0.3+1-1 | Not fixed |
| Ubuntu 22.04, 24.04, 26.04, 26.10 (USN-9155-1 to USN-9158-1) | Fixed; the *-crac packages on 26.04 and 26.10 are not |
| RHEL 8, 9 and 10 (RHSA-2026:100563 to 100572); Oracle Linux 8 and 9 | Fixed |
| CentOS Stream 9 and 10; Oracle Linux 10 | Not fixed yet |
| openSUSE Tumbleweed | Fixed. SLES 15 and SLES/Leap 16: not released yet |
| Amazon Linux 2 and 2023, Corretto 17, 21, 25, 27 (ALAS2-2026-4123, ALAS2023-2026-3478 to 3481) | Fixed |
| Alpine 3.23, 3.24, edge | Fixed. Alpine 3.22 (openjdk17 17.0.19, openjdk21 21.0.11): not fixed, and not listed in Alpine's tracker |
Docker eclipse-temurin 17.0.21_9, 21.0.13_9, 25.0.5_8 | Not fixed. The official images still ship the October builds; the update is open as adoptium/containers PR #1133 |
Docker amazoncorretto tags 17.0.21, 21.0.13, 25.0.5 | Fixed: the tag says 17.0.21, the build inside is 17.0.21.1 |
| BellSoft Liberica 17.0.21.1+5, 21.0.13.1+5, 25.0.5.1+5 | Fixed |
| Azul Zulu 17.70.201, 21.54.201, 25.38.201 | Fixed, under version strings like 17.0.20.1.201+1 that look older than the patch |
| GraalVM Community 25.1 to 25.5, and Native Image binaries built with them | Not fixed. GraalVM has its own copy of the decoder, patched in its source but in no release yet. We reproduced the write with Graal as the JIT and in a Native Image binary built with -march=native on an AVX-512 VBMI machine. A Native Image binary keeps what it was built with, whatever Java the host runs, and no JVM flag reaches it: rebuild it with a patched GraalVM |
| Oracle GraalVM for JDK 17.0.21.1, GraalVM Enterprise 21.3.20.1 | Not confirmed: Oracle's notes say these are built on 17.0.20.1+4, before the patch |
Check your own machines
# 1. Does the CPU have what the bug needs? (all four must appear)
grep -o -w -E 'avx512f|avx512bw|avx512vbmi|bmi2' /proc/cpuinfo | sort -u
# 2. Which Java build runs, and are the patched fixes in it?
java -version
# 3. What the running JVM actually uses (replace PID)
jcmd PID VM.flags -all | grep -E ' (UseAVX|UseBASE64Intrinsics|TieredStopAtLevel) '
Exposed only if step 1 shows all four flags, the build is unpatched, and step 3 shows UseAVX 3 or higher and UseBASE64Intrinsics true. A Native Image binary has no JVM flags to read: check the GraalVM it was built with.
What holds until you patch
Both flags below stopped the out-of-range write in our tests on Java 17 and 21. Neither reaches a Native Image binary, and neither replaces the patch.
| Workaround | What it does | Trade-off |
|---|---|---|
-XX:-UseBASE64Intrinsics | Turns off the fast Base64 decoders only. Also stops GraalVM's JIT path | Slower Base64 decoding. Our first choice. The flag is case-sensitive: -XX:-UseBase64Intrinsics is an unknown option and the JVM won't start |
-XX:UseAVX=2 | Makes the JVM ignore AVX-512 entirely | Turns off AVX-512 for all compiled code, not just Base64, so other work can slow down too. Unrecognised on ARM: don't push it to ARM hosts |
| Checks for "any AVX-512" | Nothing: they flag CPUs that can't be hit | Skylake-SP and Cascade Lake have AVX-512 without VBMI. Check for avx512vbmi |
| Choosing instance types by family | Not reliable | Some types land on VBMI and non-VBMI hosts; read the CPU flags on the machine |
What's happening now
- 17 November: Calif disclosed Java4Shell and Oracle shipped the patch in its November Critical Security Patch Update; OpenJDK published its advisory the same day.
- Exploitation: crashes and memory leaks have been reproduced by several research teams. Remote code execution is claimed by Calif (in Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate, with the application chains withheld) and others, but no working chain is public. Attack attempts have been seen, with no confirmed intrusions. Not in CISA's Known Exploited Vulnerabilities catalog as of its 18 November update.
- Containers lag: the official
eclipse-temurinimages still ship unpatched builds; Amazon'samazoncorrettoimages are patched. - GraalVM: its own copy of the decoder is patched in source but not in any release yet.
Updates
- 19 November 2026: published. Conditions, workarounds and the GraalVM gap confirmed on Intel Sapphire Rapids and Skylake hardware.
Sources
- Disclosure and advisories: Calif · Oracle Critical Security Patch Update, November 2026 · OpenJDK advisory · CVE record
- Distributions: Debian · Ubuntu · Red Hat · adoptium/containers #1133
- GraalVM: oracle/graal #15006 · GraalVM Community releases
- Exploitation and analysis: Tenable · Datadog Security Labs · CISA KEV