On-box check for Java4Shell, a heap out-of-bounds write in the x86-64
AVX-512 VBMI Base64 decoder intrinsic of OpenJDK HotSpot, disclosed on
2026-11-17. A malformed Base64 string with a stray = makes the decoder write
62–63 bytes past its output array before Java rejects the input. Calif
demonstrated heap disclosure and pre-auth RCE in Keycloak, Jira, Confluence,
Artifactory, Jenkins and PingFederate in their default configurations.
A host is exposed only when all of these hold:
- an affected JVM: OpenJDK 17.0.6+, 21, 25, 27 without the November 2026 fix (Oracle Java SE 8u511b33-perf and GraalVM too; standard Java 8 and Java 11 are not affected),
- an x86-64 CPU exposing
avx512f,avx512bw,avx512vbmiandbmi2(Intel Xeon Ice Lake and newer, AMD Zen 4 and newer; ARM64 is not affected), - attacker-controlled Base64 reaching
java.util.Base64(Basic auth headers, tokens, uploads, …) in a JIT-compiled decoder.
Because the CPU decides, the same container image can be vulnerable on one node and safe on another. Version lists alone are not enough, so the script runs a harmless trigger on the real runtime and CPU.
sudo bash test-java4shell.sh # whole machine incl. containers
bash test-java4shell.sh # only what the current user can see
- CPU: are all four required features visible to this OS?
- Installed runtimes (
PATH,/usr/lib/jvm,/usr/java,/opt,/usr/local,/srv,/app,/data): each one is tested by running the embeddedB64Tailclass. - Running JVMs: every process that maps
libjvm.so, also inside Docker / Podman / LXC / Kubernetes containers. The runtime the process uses is tested inside its mount namespace; then the process' effective options (JAVA_TOOL_OPTIONS,JDK_JAVA_OPTIONS, command line,_JAVA_OPTIONS, last wins) are checked for a stopgap, and the runtime is re-tested with it. The output names where the flag came from, so aMITIGATEDverdict can be checked by hand.
Verdicts: VULNERABLE / OK per runtime; for running JVMs on a vulnerable
runtime MITIGATED (stopgap active and confirmed by the re-test) or
UNPROTECTED. Example for one container:
PID 12345 (root, container /keycloak): /opt/java/openjdk/bin/java -Dkc.config.built=true ...
runtime /opt/java/openjdk/bin/java (Java 21.0.5 Eclipse_Adoptium): VULNERABLE
mitigation: -XX:-UseBASE64Intrinsics (from: command line) -> test with it OK => MITIGATED
| Exit code | Meaning |
|---|---|
| 0 | nothing vulnerable |
| 1 | a vulnerable runtime is installed, but no running JVM is exposed |
| 2 | a running JVM uses a vulnerable runtime without a stopgap |
| 3 | nothing vulnerable found, but some runtime could not be tested |
| 4 | could not create a temp directory |
The script is read-only towards services: running JVMs are never attached to
or signalled. It starts a separate short-lived JVM per runtime (128 MB heap,
about a second each) and writes the test
class to a fresh temp directory (in a container: its /tmp, /dev/shm or
/var/tmp), which is removed afterwards. All overwritten bytes are inside an
array the test class allocated itself.
- Older Java 17 builds reach the faulty code path only with 128 input
characters, not 64. A probe that only tries
64 × "A" + "="reports them as clean. The test tries 64, 128 and 256 characters with the tails=,==,A=,AA=. - Updated, not restarted: a JVM whose
libjvm.sowas replaced on disk still runs the old code. Testing the new binary would say "OK", so such a process is reported as needing a restart instead. - Inherited options: the test JVM does not inherit the caller's
JAVA_TOOL_OPTIONS/JDK_JAVA_OPTIONS/_JAVA_OPTIONS, which could otherwise hide the bug. - Untested is not clean: runtimes that could not be run are counted and turn exit code 0 into 3.
Update to the fixed build of your vendor (Oracle 17.0.21.1, 21.0.13.1,
25.0.5.1, 27.0.1.1; Temurin, Corretto, Zulu, Debian and others ship their
own), then restart the JVMs and rebuild container images. Products that
bundle a JRE need an update from their vendor.
Until then, start the JVM with one of:
-XX:-UseBASE64Intrinsics: disables only the Base64 intrinsic, works on any architecture.-XX:UseAVX=2: disables AVX-512 entirely. x86-64 only; a JVM on ARM64 does not recognise it and fails to start.
Do not use -XX:DisableIntrinsic=_base64_decodeBlock on its own; per Upwind
it stops the JVM from starting.
B64Tail.java is the source of the base64 block in the script.
javac --release 8 B64Tail.java
sha256sum B64Tail.class # put the hash in the script header
base64 -w76 B64Tail.class # replace the block between <<'B64' and B64
The current class was built with the javac of Temurin 17.0.21.1. Other
compiler versions produce different bytes and a different hash.
-
✅ Verified on an AMD EPYC 9135 (Zen 5, AVX-512 VBMI) host running Docker: Temurin 17.0.11, 21.0.2, 21.0.5 and 21.0.10 in containers and a Debian 21.0.12.1 runtime on the host were reported vulnerable, and the re-test with
-XX:-UseBASE64Intrinsicscame out clean for all of them. The Java 17 runtime had been missed by a 64-character-only probe. -
✅ Verified through a Debian package update on an AMD EPYC 9135 host, from
openjdk-2121.0.13 to the fixed 21.0.13.1 (DSA-6639-1), with two host JVMs running with-XX:-UseBASE64Intrinsics:- before the update: runtime
VULNERABLE, both JVMsMITIGATED; - after the update, before the restart: installed runtime
OK, both JVMs reported as still running the replacedlibjvm.so,MITIGATEDuntil restarted; - after the restart: both JVMs on 21.0.13.1
OK.
This confirms that a fixed build on an AVX-512 VBMI CPU reports
OK. - before the update: runtime
-
✅ Verified on a CPU without AVX-512 (AMD Zen 3) with Temurin 17.0.21.1:
OK, detection of a replacedlibjvm.soand ignoring of inherited JVM options. -
📝 Not yet verified: a Java 8 runtime, the
-XX:UseAVX=2path, Podman / LXC / Kubernetes containers.
The script only runs a self-contained trigger on the local machine. It sends nothing over the network and does not touch running services. Run it only on systems you administer.
- Calif: https://calif.io/java4shell (write-up and PoC)
- OpenJDK Vulnerability Advisory 2026-11-17: https://openjdk.org/groups/vulnerability/advisories/2026-11-17
- Upwind lab notes: https://www.upwind.io/blog/java4shell-what-we-know-now-and-what-comes-next
- Blackpoint APG analysis: https://blackpointcyber.com/blog/java4shell-cve-2026-114938-java-base64-decoder-flaw/
- Earlier public fix proposal with the trigger: openjdk/jdk#32781
MIT, see LICENSE.