Java4Shell: What to Do Before Your JDK Patch Lands
Java4Shell (CVE-2026-114938) has a fix, but four of five JDK images I read on Nov 18 lacked it. Triage: CPU check, find JVMs, set a flag, restart, patch.
In this post
TL;DR: Java4Shell (CVE-2026-114938) only affects x86-64 hosts with AVX-512 VBMI, according to Calif, which reported it. Check the CPU, find every JVM including bundled ones, set Calif’s mitigation flag
-XX:-UseBASE64Intrinsics, and restart. Then patch when your vendor ships. At about 19:33 UTC on November 18, Temurin 17, 21 and 25 had no fixed release, and the Keycloak and Jenkins image tags I read (ARM64 variants) still shipped unfixed JDKs.
On Tuesday, November 17, 2026, Oracle released a patch and the OpenJDK project published an advisory for CVE-2026-114938, a Base64 bug in HotSpot that the security firm Calif named Java4Shell. Calif says both went out at 1:00pm PT, which is 21:00 UTC.
At about 19:33 UTC on November 18, Adoptium’s release API still listed unfixed Temurin 17, 21 and 25 builds for Linux x64, and four of the five common image tags I read (the ARM64 variants) shipped an unfixed JDK. So decide who is exposed, set one flag, restart, and patch when your vendor ships. Every “as of” line below comes from pages I read on November 18. The vendor lists were changing by the hour, so check them again before you act. This follows the format of my triage posts for the PHP flaws fixed on October 21 and the OpenSSH sshhole bug on Alpine.
Calif’s write-up covers the mechanism. In short: the bug is in the
x86-64 Base64 decode intrinsic, a hand-written assembly stub inside
HotSpot, not the Java-level java.util.Base64 code. For a
padded input tail, the stub miscalculates its store mask and writes past
the end of the output array.
The three severity scores
At least three scores are going around, from three different sources:
- Oracle, the CNA that assigned the CVE, rates it CVSS 3.1 8.1 High, with attack complexity High. The OpenJDK advisory also says 8.1.
- Calif rates it CVSS 4.0 9.2 Critical, on its own page.
- Red Hat’s CVE page said 9.8 and called the score preliminary.
NVD showed the record as “Undergoing Analysis” in an API response stamped 17:37 UTC on November 18. CISA’s alert, titled “Critical Java Vulnerability Requires Immediate Patching”, describes it this way: “Specially crafted Base64 input could crash the JVM and allow for unauthenticated remote code execution.”
Oracle’s and Red Hat’s scores differ on one field. Oracle rates attack complexity High. Red Hat’s preliminary score rates it Low. Both say the attack comes over the network with no login. Calif’s page backs the High rating: it says there is “no universal exploit” and that each target needs its own profile. Per Tenable, VulnCheck does not expect the flaw to be easily or widely exploitable, but says crashing an affected application is easy. I still would not use the 8.1 to put off the flag and restart. Calif says most of its later chains were built “almost autonomously with AI”, and a crash takes much less work than code execution.
Who is exposed
All of this comes from Calif’s page unless I say otherwise.
- CPU: x86-64 with AVX-512 VBMI. The host needs all four flags: avx512f, avx512bw, avx512vbmi and bmi2. Calif lists Intel Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD Zen 4 or newer. ARM64 is not affected.
- JDK: OpenJDK 17.0.6 or newer, across the vendor builds. Calif names Corretto, Temurin, Microsoft, Zulu, Liberica, SapMachine, Red Hat, Oracle and GraalVM builds.
- Lines: the OpenJDK advisory lists 17, 21, 25 and 27, and not 8 or 11. Oracle’s November patch page lists one Java 8 build for this CVE, 8u511b33-perf, which Tenable says is the Enterprise Performance Pack. Standard Java 8 and 11 are not listed for it. Per Tenable, Oracle lists only builds it supports and says earlier versions are likely also affected. Aikido’s table lists the end-of-life lines 18 through 24 and 26 as affected, with an upgrade to a fixed 21, 25 or 27 build as the target.
- JIT: the method must be warm enough that the JIT picks the intrinsic. Calif’s trigger runs 200,000 valid decodes first.
Calif says it demonstrated pre-authentication remote code execution in default configs on Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate. The evidence is a demo video I did not watch, and Calif says “The weaponized chains for individual products stay private until everyone has had a chance to upgrade.”
Calif also lists products it exploited after login: Kafka, Solr, Elasticsearch, Gerrit and Spinnaker. It says it ran code in sample apps built on Struts and Spring Boot, and that a Java program that opens a TLS connection to a malicious server can be compromised even if it exposes no service of its own. So build agents and batch jobs that connect out are in scope too.
Is it being exploited?
As of 19:40 UTC on November 18, I found no source reporting a confirmed compromise. Attempts and private exploits were being reported by then.
| Source | What it says |
|---|---|
| CyberScoop, Nov 17 | “No exploitation in the wild had been reported as of this article’s publication” |
| Aikido, Nov 18 | No confirmed exploitation. “Other researchers have so far reproduced crashes and memory leaks rather than code execution.” |
| NVD record | CISA’s SSVC entry, stamped 21:53 UTC Nov 17: exploitation none |
| Tenable, updated Nov 18 | “GreyNoise has reported exploitation attempts against CVE-2026-114938, though GreyNoise has not confirmed any successful exploitation against its sensors.” |
Aikido’s line is older than Tenable’s update. Tenable says VulnCheck has a working code-execution exploit available to its customers, and that Kevin Breen of Immersive said on X he has a working exploit. Tenable notes his post does not say it has been released. Per Tenable, Breen said Apache Tomcat’s Basic authentication and the java-jwt library give a route before login. I read GreyNoise, VulnCheck and Breen only through Tenable.
Calif’s public proof of concept only triggers memory corruption and runs no code. Calif warns that failed attempts may crash the JVM, so unexplained restarts are worth a log check.
Triage, in order
1. Find the hosts with the CPU
This is Calif’s check. Run it on every node that serves production if you can, because the flags depend on the host CPU generation. Some cloud instance families mix generations, so one clean result does not clear the family:
grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' \
| grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'All four flags must print for the host to be affected. On my ARM64 Linux laptop, in a container, it printed nothing and grep exited 1, the expected result on ARM64.
2. Find every JVM, running and on disk
I ran every command in this post in Docker containers on that ARM64
laptop between 19:32 and 19:34 UTC on November 18. That covers the
inventory, flag and restart checks. I did not reproduce the bug, because
ARM64 is not affected. Calif’s published trigger printed
NOT REPRODUCED on unfixed Temurin 21.0.13+9 there, which
says nothing about x86-64.
This lists running JVMs without ps. Run it as root:
grep -l libjvm.so /proc/[0-9]*/maps 2>/dev/nullIt found both of my test JVMs. For each PID, get the binary, then run
that path with -version:
readlink /proc/<pid>/exe
/opt/java/openjdk/bin/java -versionThe first printed /opt/java/openjdk/bin/java, the second
Temurin-21.0.13+9. Then find every JVM on disk, including
copies bundled inside an app, which a system update misses. Each
libjvm.so sits in lib/server/ under its JDK
directory, and the release file at the top of that
directory has the build:
find / -xdev -name libjvm.so 2>/dev/null
grep -E '^(IMPLEMENTOR|JAVA_RUNTIME_VERSION)=' \
/opt/java/openjdk/releaseIn the Temurin image that gave
JAVA_RUNTIME_VERSION="21.0.13+9-LTS". The
-xdev flag stays on one filesystem, so run it again against
mounted volumes. The amazoncorretto:21 image has no
find, but java -version worked there.
3. Rank what you found
This ordering is mine. Calif’s product lists are the input.
- Internet-facing Java that an unauthenticated caller can send Base64 to, on an AVX-512 VBMI host, on JDK 17.0.6 or later. Calif says it has pre-authentication chains for Keycloak, PingFederate, Jenkins, Artifactory, Jira and Confluence. Your own services belong here too if they handle a login or an upload before authentication. Calif says “Ordinary features such as authentication and file uploads can expose the vulnerable Base64 decoder”. I’d start with the identity provider, because everything behind single sign-on depends on it.
- Everything else on an affected CPU and JDK line: internal services, the products Calif exploited after login, and Java clients that connect out to servers you don’t control.
- No action for this CVE: ARM64 hosts, which Calif says are not affected, and standard Java 8 and 11 builds, which the OpenJDK advisory does not list.
Don’t count on a reverse proxy’s request buffering. Calif says it blocks some of its exploit profiles and not others. The proxy has its own patch this week, covered near the end.
4. Set the flag, then restart
Calif lists two mitigations. -XX:-UseBASE64Intrinsics
turns off hardware-accelerated Base64. -XX:UseAVX=2 turns
off AVX-512 and exists only on x86-64. I have no performance number for
either, and could not test whether either stops the bug without an
affected CPU. Per Tenable, neither Oracle nor the OpenJDK Vulnerability
Group mentions these options in its advisory. On November 18 the
Keycloak project recommended -XX:-UseBASE64Intrinsics as an
interim mitigation.
I could test whether the flags are accepted. On Temurin 21.0.13:
java -XX:+PrintFlagsFinal -version \
| grep UseBASE64IntrinsicsThat printed
bool UseBASE64Intrinsics = true {product} {default}, and
= false {product} {command line} with
-XX:-UseBASE64Intrinsics added. So ARM64 accepts the flag
too.
On ARM64, -XX:UseAVX=2 stopped the JVM from starting. It
printed Unrecognized VM option 'UseAVX=2' followed by
Error: Could not create the Java Virtual Machine. Calif
warns about this. In a mixed x86 and ARM fleet, a blanket
UseAVX=2 takes the ARM services down.
-XX:-UseBASE64Intrinsics is safe on both.
Run the PrintFlagsFinal check on each JDK build you have
before you roll the flag out. A JVM that does not recognize an option
will not start.
To set it without editing a start script, put
JAVA_TOOL_OPTIONS=-XX:-UseBASE64Intrinsics in the service
environment. The JVM printed
Picked up JAVA_TOOL_OPTIONS: -XX:-UseBASE64Intrinsics at
start.
A running JVM keeps the flags it started with, so restart it. Then
confirm with jcmd (JDK only):
jcmd <pid> VM.flags | tr ' ' '\n' | grep -i base64On the mitigated process that printed
-XX:-UseBASE64Intrinsics. On the unmitigated one it printed
nothing, because VM.flags lists only non-default flags.
jcmd <pid> VM.flags -all | grep UseBASE64 printed the
value on both.
A JRE image has no jcmd. There,
tr '\0' '\n' < /proc/<pid>/environ | grep JAVA_TOOL_OPTIONS
showed the variable on the mitigated process and nothing on the
other.
CISA also suggests validating Base64 input at the application boundary, input-size limits and limiting unauthenticated access. I would add those where the flag can’t be set yet.
5. Patch when your vendor ships, and read the version string
Oracle’s fixed builds, from Tenable’s table, are 8u513-perf, 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. For Oracle, Temurin and Corretto builds, a 17.0.21, 21.0.13 or 25.0.5 version without a 1 in the fourth position is unfixed. Anything older is unfixed too. Tenable notes that other distributions number their builds differently.
A JVM on an end-of-life line has no fixed build in Aikido’s table. Set the flag and move it to 21, 25 or 27.
This is what the common images shipped at 19:32 UTC on November 18, and again when I re-read them at 20:27 UTC. I read the ARM64 variant of each tag, not the x86-64 ones:
| Image tag | JDK inside | Fixed build? |
|---|---|---|
eclipse-temurin:21-jdk |
Temurin 21.0.13+9 | No |
eclipse-temurin:17-jre |
Temurin 17.0.21+9 | No |
amazoncorretto:21 |
Corretto 21.0.13.1.2 | Yes |
quay.io/keycloak/keycloak:latest |
Red Hat build 21.0.13+9 (21.0.13.0.9-1) | No |
jenkins/jenkins:lts-jdk21 |
Temurin 21.0.13+9 | No |
Adoptium’s
release API, read at 19:33 UTC and again at 20:27 UTC, returned
jdk-17.0.21+9, jdk-21.0.13+9,
jdk-25.0.5+8 and jdk-27.0.1.1+4 for Linux x64,
so only Temurin 27 was fixed. Corretto’s changelogs for 17,
21
and 25,
read around 19:35 UTC, list the CVE under 17.0.21.1.2, 21.0.13.1.2 and
25.0.5.1.2. The Corretto 27 changelog
shows 27.0.1.1.4.
For other vendors I have less. Tenable’s November 18 update says “Amazon Corretto and Azul Zulu have released fixed builds, and Eclipse Temurin has begun releasing them,” and that Red Hat and Ubuntu had not released fixes. Red Hat’s CVE page said Oracle had not yet confirmed the fixed version or the affected release lines. I did not check Atlassian, Jenkins or Ping.
After the upgrade, restart again. This lists JVMs running a
libjvm.so that was replaced on disk:
grep -l 'libjvm.so (deleted)' \
/proc/[0-9]*/maps 2>/dev/nullI replaced libjvm.so under two running JVMs to test it.
It printed nothing before and both processes after.
This check is for package upgrades on a host or VM. A container gets the fix from an image rebuilt on a fixed base tag, so check the JDK version inside the image each service runs.
What to tell your board or a customer
The paragraph I’d start from:
On November 17 Oracle released a fix and the OpenJDK project published an advisory for CVE-2026-114938, a Java vulnerability that CISA says could allow unauthenticated remote code execution. According to Calif, which reported the bug, only x86-64 servers with AVX-512 VBMI processors are affected. We identified [N] JVMs on [N] hosts, of which [N] run on affected processors. [N] have the vendor’s fixed build as of [date and time]. [N] have Calif’s published mitigation (
-XX:-UseBASE64Intrinsics) set and were restarted, pending a fixed build expected [date or “not announced”]. As of [date and time], [source] reports no confirmed exploitation, and we are reviewing logs for unexpected JVM restarts. We will confirm completion on [date].
The same customer may send a full security questionnaire next. I wrote up how those reviews go for AI startups.
What Calif says it cost to find
Calif says it took “an hour to discover the vulnerability, and two more hours to develop the first working exploit”, at a token cost of “less than $200”. Those are Calif’s own numbers. I found no independent confirmation.
Calif also says Quang Luong spent around 20 minutes prompting past junk results before Codex produced the finding, and that it built pipelines and strategies first to speed up exploit development.
The most concrete number in my post on old bugs that AI now finds in a week was three days and under $700 for PwnEscape. Calif’s claim is lower. In both cases people steered the work. What I’d plan for is the list this post started with: every JVM, the CPU it runs on, the vendor that ships its JDK, and when that vendor’s fix arrives. The restart and patch labor after that is the fix-side constraint for small teams I wrote about in July.
If nginx fronts your Java services
If nginx fronts your Java services, it goes on the same patch list, starting where HTTP/2 is on. On November 18, the day after Java4Shell, nginx released 1.31.9 (mainline) and 1.30.7 (stable). Xint published five of the fixed CVEs as nginx2root. Per Xint’s write-up, they give code execution as the nginx worker process, and reaching root uses a separate Linux kernel bug. Per Xint’s table, each affects a specific configuration (HTTP/2, HTTP/3, mail proxy, or resolver plus mirror), and Xint says the HTTP/2 one reaches the most servers. At about 19:25 UTC on the 18th, Debian’s tracker still listed its nginx packages as vulnerable.
I do fractional security work for startups. If you want a second person on triage like this, or on the customer answers afterward, the engagements are on my consulting page.