Java4Shell (CVE-2026-114938): check your JVM and CPU
Java4Shell is an OpenJDK Base64 heap bug on x86-64 with AVX-512 VBMI. Check your CPU flags, java -version and bundled JVMs, then patch or set the stopgap.
Quick summary · TL;DR
- Run the CPU flag check before anything else. If
/proc/cpuinfolacks any of avx512f, avx512bw, avx512vbmi or bmi2, Java4Shell can't fire on that box today, and every ARM64 server is out of range. - Then compare
java -versionwith 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. Anything older on 17.0.6 or later is in range. Java 8 and 11 don't have the vulnerable code, per Red Hat. - Images carry their own JVM, so check inside them.
docker run --rm --entrypoint java IMAGE -versiontakes seconds, and Red Hat already lists its Keycloak and Kafka images as affected. - Can't patch today? Set
-XX:-UseBASE64Intrinsics. Calif's other flag,-XX:UseAVX=2, stops the JVM starting on ARM, which is easy to miss on a multi-arch image.
If any JVM you run sits on an x86-64 server with AVX-512, patch it today. Java4Shell (CVE-2026-114938) is a heap memory bug in OpenJDK’s Base64 decoder, fixed on 17 November 2026. It needs two things at once: JDK 17.0.6 or later, 21, 25 or 27, and a CPU with AVX-512 VBMI. Fixed builds are 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1.
Both checks take a minute. Run the CPU flag one-liner and java -version below, on the host and inside every image.
What Java4Shell is
The OpenJDK advisory of 17 November 2026 lists CVE-2026-114938 under hotspot/compiler at CVSS 3.1 8.1 and marks JDK 17, 21, 25 and 27 as affected. The CVE record, filed by Oracle, describes an unauthenticated attacker with network access who can take over the Java runtime, in a flaw rated difficult to exploit. Oracle shipped the fix in its November 2026 Critical Security Patch Update.
Calif reported the bug, named it and published the technical write-up. The bug sits in the x86-64 Base64 intrinsic, hand-written machine code the JIT uses to decode faster. According to Calif’s write-up, it miscalculates the tail of padded input and writes past the end of the output array. Its example trigger is 64 A characters followed by one = (128 on older JDK 17 builds). The fix commit adds four lines that hand short padded tails back to the plain Java code.
Java4Shell severity: how bad is it
| Source | Score | What it says |
|---|---|---|
| Oracle, OpenJDK | CVSS 3.1 8.1 | High, attack complexity high, no login needed |
| Red Hat | CVSS 3.0 9.8 | Critical, cites public research with working exploits |
| Calif | CVSS 4.0 9.2 | Critical |
Calif’s tests got remote code execution in two groups of apps.
- Pre-auth RCE in default configs of Keycloak, Atlassian Jira and Confluence, JFrog Artifactory, Jenkins and PingFederate.
- Post-auth RCE in Kafka, Solr, Elasticsearch, Gerrit and Spinnaker.
By Calif’s account there’s no universal exploit. Each app needs its own chain, and those chains stay private for now. CISA’s entry in the CVE record (17 November) lists exploitation as none.
A misplaced write can crash the JVM, and Calif says an attacker can simply try again. Until a host is patched, treat an unexplained crash there as a lead. HotSpot writes an hs_err_pid<pid>.log file to the process’s working directory by default, per Oracle’s troubleshooting guide. CISA’s watch list adds Java runtimes nobody accounted for and Java processes doing things they shouldn’t.
Who is affected by Java4Shell
Red Hat’s statement names both conditions: an x86-64 CPU with the avx512f, avx512bw, avx512vbmi and bmi2 features, and an OpenJDK build from 17.0.6 on. ARM64 is not affected, and Red Hat says OpenJDK 8 and 11 don’t contain the vulnerable code.
| JDK line | Affected | Fixed in |
|---|---|---|
| 8, 11 | No (Oracle lists only its own 8u511b33-perf build) | n/a |
| 17 | 17.0.6 to 17.0.21.0.3 | 17.0.21.1 |
| 21 | up to 21.0.13.0.3 | 21.0.13.1 |
| 25 | up to 25.0.5.0.3 | 25.0.5.1 |
| 27 | up to 27.0.1.0.3 | 27.0.1.1 |
Debian, Ubuntu and Red Hat ship these fixed numbers. Vendor builds add their own suffix, so Corretto’s is 27.0.1.1.4, per its changelog. Oracle’s CVE text also lists Oracle GraalVM for JDK 17.0.21.0.3 and 21.0.13.0.3 and GraalVM Enterprise Edition 21.3.20.0.3 as affected.
Check your systems
1. The CPU flags
This is Calif’s command. All four flags have to show up:
grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' | grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'
If you’d rather see which ones are missing, this prints a yes or no per flag:
for f in avx512f avx512bw avx512vbmi bmi2; do grep -qw "$f" /proc/cpuinfo && echo "$f yes" || echo "$f no"; done
Calif names Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, plus AMD EPYC and Ryzen from Zen 4, as CPUs that qualify. Containers see the host’s flags, so run it on the node. Treat a “no” on a cloud VM as today’s answer only, because a resize or a new instance type can land you on an AVX-512 CPU without a code change.
2. Every JVM, not only the system one
java -version on the host is the easy part. Images and apps that bring their own runtime are where old builds hide.
java -version
docker run --rm --entrypoint java IMAGE -version
kubectl exec deploy/NAME -- java -version
To find runtimes bundled inside apps, look for java binaries and JDK release files:
find / -type f -name java -perm -u+x 2>/dev/null
find / -name release -path '*j*' -exec grep -H JAVA_VERSION {} \; 2>/dev/null
Red Hat lists its own Keycloak image, a ubi10 OpenJDK 21 image and several Kafka and EAP images as affected, and patching the host package fixes none of them. Rebuild each image on a fixed base and check the tag again.
3. Calif’s safe test
Calif’s write-up includes a short, self-contained Java program. It runs about 200,000 valid decodes so the JIT switches to the vulnerable fast path, then decodes the 64 A plus = string into a 4,096-byte array and prints REPRODUCED on an affected JDK. Every write stays inside a real array, Calif says, so it’s safe to run. One decode in jshell won’t tell you anything, since the bug only lives in the compiled path. Run Calif’s program with the same JVM and on the same host you care about.
Fix Java4Shell, or set the stopgap
Upgrade to the fixed build for your line, then restart every JVM. A running process keeps the old runtime loaded until it exits. CISA’s alert of 17 November says to patch internet-facing systems first, and Calif’s pre-auth list (Keycloak, Jira, Jenkins and the rest) tells you which of those to start with.
If you can’t patch yet, there are two flags in Calif’s write-up:
-XX:-UseBASE64Intrinsicsturns off hardware-accelerated Base64 encoding and decoding.-XX:UseAVX=2turns off AVX-512, on x86-64 hosts only.
For containers, set the flag through the environment:
ENV JAVA_TOOL_OPTIONS="-XX:-UseBASE64Intrinsics"
JAVA_TOOL_OPTIONS is read by the JVM itself, and the JVM TI spec describes it for VMs launched deep inside scripts or embedded. JDK_JAVA_OPTIONS works too, but only for the java launcher, per the java command docs. If the app already sets JAVA_TOOL_OPTIONS, append to it rather than replace it. Then confirm it took:
java -XX:+PrintFlagsFinal -version | grep UseBASE64Intrinsics
You want false.
Request buffering, on by default in nginx, breaks some of Calif’s chains, but the write-up calls it no general mitigation. It doesn’t stop the Elasticsearch chain, and WebSocket frames skip it.
Distro packages
| Distro | Status on 20 Nov 2026 |
|---|---|
| Debian bookworm | openjdk-17 fixed in 17.0.21.1+2-1~deb12u1 (DLA-4917-1) |
| Debian trixie | openjdk-21 21.0.13.1+2-1 |
| Debian forky | Still listed as vulnerable on the Debian tracker |
| Ubuntu 22.04 to 26.10 | openjdk-17, 21 and 25 fixed (USN-9155-1 and follow-ups), see the Ubuntu CVE page |
| Ubuntu 18.04, 20.04 | Fixed through ESM (Ubuntu Pro) |
| RHEL 8, 9, 10 | Fixed on 19 November: java-17 and 21 on RHEL 8, java-17, 21 and 25 on RHEL 9, java-21 and 25 on RHEL 10 (e.g. RHSA-2026:100567) |
Compare dpkg -l 'openjdk*' or rpm -qa 'java-*-openjdk*' with the tracker for your release.
Why scraping stacks should care about Java4Shell
There’s a client side too. Calif says samczsun shared an exploit where a Java program that opens a TLS connection to a malicious server can be compromised even with no service of its own. A Java crawler connects to servers it doesn’t control all day. Routing it through rotating residential proxies changes the IP a target sees. It does nothing to what that server sends back, so the JVM patch is the fix. Servers around a web scraping architecture need the same pass, as with the nginx HTTP/2 bugs yesterday.
What to do now
Run the CPU check on every x86-64 host, then java -version everywhere a JVM lives. Patch the exposed ones to 17.0.21.1, 21.0.13.1, 25.0.5.1 or 27.0.1.1 and restart them. Anything you can’t patch today gets -XX:-UseBASE64Intrinsics.
Frequently asked questions
Java4Shell is the name Calif gave CVE-2026-114938, a heap memory corruption bug in OpenJDK's x86-64 Base64 decoder intrinsic. Malformed Base64 input can make it write past the end of the output array. Oracle and OpenJDK fixed it on 17 November 2026 and rate it CVSS 3.1 8.1.
OpenJDK 17 from 17.0.6, plus 21, 25 and 27, up to 17.0.21.0.3, 21.0.13.0.3, 25.0.5.0.3 and 27.0.1.0.3. It also needs an x86-64 CPU with avx512f, avx512bw, avx512vbmi and bmi2. Builds from vendors such as Temurin, Corretto and Zulu share the same code, per Calif.
17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1, or later. Those are the numbers Debian, Ubuntu and Red Hat ship. Vendor builds add their own suffix, for example Corretto 27.0.1.1.4.
No for OpenJDK. Red Hat says OpenJDK 8 and 11 don't contain the vulnerable code path, and the OpenJDK advisory marks only 17, 21, 25 and 27. Oracle's CVE text does list its own 8u511b33-perf build, so check that one if you run it.
On Linux, run Calif's command: grep -m1 '^flags' /proc/cpuinfo, split it into lines and look for avx512f, avx512bw, avx512vbmi and bmi2. If all four appear, the CPU qualifies. ARM64 servers are not affected.
Start the JVM with -XX:-UseBASE64Intrinsics, for example through JAVA_TOOL_OPTIONS in a container. Calif also lists -XX:UseAVX=2, but OpenJDK on ARM doesn't recognize that flag and won't start with it. Upgrade as soon as you can.
CISA's entry in the CVE record listed exploitation as none on 17 November 2026. Calif says it built working exploits for apps like Jenkins and Keycloak but is holding the app-specific chains back for now.
Researched and drafted with AI tools, then checked against our publishing rules before it went live: 13 outside links, 3 links to other proxymint pages, no prices and a fact-check pass against our product records. How we write.
