Java4Shell (CVE-2026-114938) is a reported remote code execution flaw in Java itself, not in a library: a heap buffer overflow in the hardware-accelerated Base64 decoder that Java uses on x86 servers with AVX-512. Because Base64 shows up in login headers, tokens and uploads, the vulnerable code can run before anyone logs in. Researchers at Calif found and reported it; Oracle credits them for it in its 17 November 2026 security update, which ships fixed builds such as 17.0.21.1 and 21.0.13.1. If you run Java 17.0.6 or later on a recent Intel or AMD server, update the JDK, restart every JVM and rebuild your container images. If you can't patch today, a JVM flag that switches off the vulnerable decoder path is the stopgap most researchers point to.
I'm Kunal Vohra, a fractional CTO and a security researcher and cybersecurity advisor. Most of the fix guides on this blog have been for the JavaScript and Linux side of a startup stack, like the Next.js CVE-2026-96889 upgrade and the APT fix. This one matters even if your product is not written in Java, because the tools around your product often are: Jenkins, Keycloak, Jira, Confluence, Artifactory, Kafka, Elasticsearch. Here is what is confirmed, what is still only reported, and how to close it.
What Java4Shell is, and how sure we are
A lot of the detail is still coming from researchers and security vendors rather than from a full official advisory, so I'll separate the two.
Confirmed by Oracle. Oracle published a Critical Security Patch Update on 17 November 2026 (Rev 1, "Initial Release"). Its Java SE entry lists 8u511, 11.0.33, 17.0.21, 21.0.13, 25.0.5 and 27.0.1 among the affected versions, and its credits name "Quang Luong of Calif.io" and Albert Yang for CVE-2026-114938. Security vendors including Qualys, Aikido and Cycode report that Oracle's highest Java SE score in this update is 8.1 under CVSS 3.1.
Reported by researchers and vendors, not yet in an official write-up I could read.
- The component. Qualys, Orca, Oligo and Cycode describe the bug as an integer underflow in padding handling inside the AVX-512 Base64 decoding intrinsic (the hand-written assembly HotSpot uses to speed up
java.util.Base64). A crafted 65-byte input reportedly writes 62 bytes past the end of a heap buffer. Aikido and Endor Labs say the component has not been officially disclosed. - The version range. Datadog Security Labs, Qualys, Orca and Cycode say OpenJDK-based builds from 17.0.6 onwards. Qualys and Cycode include 21, 25 and 27; Orca calls those likely but unconfirmed; Endor says only the Java 17 line has been publicly tied to this CVE. Standard Java 8 and 11 are reported as not affected, even though Oracle shipped new builds for them in the same update (that update fixes other issues too).
- The CPU. Only x86 servers with AVX-512 are reported as exposed. Qualys and Cycode list four CPU flags that must all be present:
avx512f,avx512bw,avx512vbmiandbmi2. That means Intel Ice Lake and later server chips and AMD Zen 4 and later. ARM servers, including AWS Graviton, are reported as not affected. - The score. 9.2 under CVSS 4.0, according to the researchers (Qualys) and early reporting, with no published vector. Oracle's 8.1 may or may not describe the same issue; Aikido says the two figures "may describe different issues".
- Exploitation. As of 17 to 18 November, Qualys, Aikido, Orca and Endor all say there is no confirmed exploitation in the wild and no public proof of concept, and it is not in CISA's KEV catalogue. Qualys reports that the researchers built working exploits for several products and are holding them back, and that they found the bug with AI assistance in about an hour.
That last point is why I'm not waiting for the paperwork. I wrote yesterday about Google's exploit_agent report and AI turning patches into exploits; a public patch plus a public description of the root cause is exactly the input those tools need.
The line I'd put in front of any founder: your product might not be written in Java, but your build server, your login server and your issue tracker probably are.
Is this you? Who should care
You are in scope if all three are true:
- You run a Java runtime from 17.0.6 onwards that predates the November fixes. That includes Oracle JDK, OpenJDK, Eclipse Temurin, Amazon Corretto, Azul Zulu, Red Hat builds, distro packages, and Java bundled inside products and Docker images.
- It runs on x86 hardware with AVX-512. Most current-generation cloud instances on Intel Ice Lake or later and AMD Zen 4 or later qualify. Graviton and other ARM machines are reported as safe from this specific bug.
- Something reachable decodes attacker-controlled Base64. Cycode says the trigger is input passed to the basic or URL decoder of
java.util.Base64. In practice: HTTP Basic auth headers, tokens, encoded API fields and file uploads.
Qualys published a table, credited to Calif, of products where the researchers demonstrated code execution in default configurations. Without logging in: Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate. With a login: Apache Kafka brokers, Apache Solr, Elasticsearch, Gerrit and Netflix Spinnaker. Qualys stresses that the list is not exhaustive and mentions sample Spring Boot and Struts apps. It also reports a client-side angle: a Java program can reportedly be compromised by connecting over TLS to a malicious server, so outbound connections matter too, not only what you expose.
If you use the cloud versions of Jira, Confluence or Artifactory, patching the runtime is the vendor's job; watch their status pages. If you self-host any of them, it's yours.
The five-minute check
Run these on each Linux host that runs Java. They only read information; none of them change anything.
1. Which Java is installed?
java -version 2>&1
2. Does the CPU have the four flags? (Command as published by Qualys and Cycode. Exposed only if all four lines print.)
grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' | grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'
3. Is the Base64 intrinsic switched on in this JVM?
java -XX:+PrintFlagsFinal -version 2>/dev/null | grep -E ' (UseAVX|UseBASE64Intrinsics) '
UseBASE64Intrinsics set to true and UseAVX set to 3 on an affected version and CPU is the combination the researchers describe.
4. Which Java is each running process actually using? The java on your PATH is often not the one your service runs.
ps -eo pid,args | grep '[j]ava'
sudo readlink -f /proc/<pid>/exe
sudo jcmd <pid> VM.version
Run jcmd as the same user as the process if it refuses to attach.
5. Containers. Images carry their own JDK, and patching the host does nothing for them.
docker run --rm --entrypoint java <image> -version
On Kubernetes, list the images you run so you can check each Java-based one:
kubectl get pods -A -o jsonpath='{range .items[*]}{.spec.containers[*].image}{"\n"}{end}' | sort -u
How to fix it, step by step
Step 1: Install a fixed JDK from your vendor
Oracle JDK. According to Qualys, Aikido, Cycode and Orca, the fixed builds released on 17 November are 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1, plus 11.0.33.1 and 8u513 (8u513-perf for the Enterprise Performance Pack). Oracle's advisory lists the affected versions; check its release notes for the build you download.
Amazon Corretto. Qualys reported on 17 November that Amazon had released Corretto 17.0.21.1.2 with the fix. Confirm in the Corretto changelog for your major version.
Eclipse Temurin, Azul Zulu, Red Hat, Microsoft and distro packages. None of the sources I read gave fixed build numbers for these yet; as Qualys puts it, other OpenJDK vendors "publish on their own schedules." Check your vendor's advisory page today, and again tomorrow. On Debian and Ubuntu, see what you have and whether an update is waiting:
apt list --installed 2>/dev/null | grep -i openjdk
sudo apt-get update && apt list --upgradable 2>/dev/null | grep -i openjdk
When your distro's security tracker lists a fixed package for CVE-2026-114938, upgrade only that:
sudo apt-get install --only-upgrade openjdk-17-jre-headless
On RHEL-family systems, sudo dnf check-update 'java-*-openjdk*' shows what is pending.
Step 2: Restart every JVM
A new JDK on disk does nothing for a process that started last week. Restart each service and confirm with jcmd <pid> VM.version that it picked up the new build. This is the step people skip.
Step 3: Rebuild and redeploy images
Pull a fixed base image and rebuild without the cache, so you don't reuse an old layer:
docker build --pull --no-cache -t myapp:java4shell-fix .
If your Dockerfile pins a tag like eclipse-temurin:17.0.20_8-jre, move to the fixed tag once your vendor publishes it. A floating tag like eclipse-temurin:17-jre only helps after the vendor has pushed the new image and you have rebuilt.
Step 4: If you can't patch today, switch off the vulnerable path
Qualys's preferred stopgap is to disable the Base64 intrinsics, so Java falls back to its standard decoder. You can apply it without touching the app's start script through JAVA_TOOL_OPTIONS:
# systemd: sudo systemctl edit myservice, then add:
[Service]
Environment="JAVA_TOOL_OPTIONS=-XX:-UseBASE64Intrinsics"
# Docker
docker run -e JAVA_TOOL_OPTIONS="-XX:-UseBASE64Intrinsics" myapp
Restart and confirm with step 3 of the check that UseBASE64Intrinsics now reads false. Three cautions. First, no vendor has validated this as a fix; Orca says plainly that "No vendor-validated workaround has been published," and Qualys says "Neither option replaces the patch." Second, the alternative some sources mention, -XX:UseAVX=2, turns off more optimisations and is x86-only: Cycode warns OpenJDK on ARM "does not recognize it and fails to start," so don't push it to a mixed fleet. Third, Base64 decoding gets slower, which you may notice on services that decode heavily; measure before and after.
Step 5: Put the exposed tools behind something
While you wait on vendor builds, reduce who can reach the Java services on the list above. Jenkins, Artifactory, Kafka, Solr and Elasticsearch rarely need to be on the public internet at all; put them behind a VPN, an identity-aware proxy or an IP allowlist. Qualys notes that reverse-proxy buffering is "not a general mitigation," so a proxy in front is about access, not about filtering the payload.
Step 6: Look for signs of trouble
There is no public exploit yet, but if a service was exposed before you patched, check for JVM crash files and anything odd the Java process started:
sudo find / -xdev -name 'hs_err_pid*.log' -mtime -7 2>/dev/null
sudo journalctl --since "7 days ago" | grep -iE 'SIGSEGV|hs_err' | head
ps -eo pid,ppid,user,args --forest | grep -A3 '[j]ava'
A crash log from a public-facing service this week is worth a closer look; a shell or curl running as a child of your Java process is an incident. If you find either, preserve the logs, isolate the host and rotate the secrets that service could read.
What it means for startups, and what I'd do as a CTO
It's easy for a team to say "we're a Node shop" or "we're a Python shop" and stop reading. That's the trap. Java tends to live in the infrastructure nobody thinks of as product: the self-hosted Jenkins that holds deploy keys, the Keycloak that issues every login, the Kafka cluster, the Elasticsearch behind search, the Atlassian server someone set up in year one. The researchers' list reads like the back office of a typical Series A company, and two of those products (Jenkins and Keycloak) hold exactly the secrets an attacker wants.
If I were CTO at a startup today, I'd do this in order:
- Inventory first, in an hour. Every host and image with a JVM, which vendor, which version, which CPU family. Search Dockerfiles for
openjdk,temurin,correttoandzulu, and remember the products that bundle their own Java. - Patch the pre-auth list first. Anything on the "without logging in" list that is reachable from the internet gets the stopgap flag today and a vendor build as soon as it exists.
- Take build and identity tools off the internet. Jenkins and Keycloak should not be one header away from code execution even on a good day.
- Make the next one boring. Pin base images by version, rebuild them on a schedule, and subscribe to your JDK vendor's advisories and CISA's KEV feed. My startup security checklist covers that routine, and the Node.js security release guide is the same pattern for the JavaScript side.
The common lesson from every fix guide on this blog: the patch is quick; finding every place that needs it is the slow part. If nobody on your team owns that inventory, a technical advisor can set it up with your engineers in a few sessions, and a fractional CTO owns it on an ongoing basis along with the rest of your security posture.
Frequently asked questions
What is Java4Shell? Java4Shell is the name researchers gave CVE-2026-114938, a reported heap buffer overflow in the AVX-512 Base64 decoder built into the Java runtime (HotSpot) from 17.0.6 onwards. Because applications decode Base64 from login headers, tokens and uploads, it can reportedly lead to remote code execution without authentication on x86 servers with AVX-512.
Is Java 17 affected by CVE-2026-114938? Yes, Java 17.0.6 and later builds before the November 2026 fixes are reported as affected on x86 CPUs that have the avx512f, avx512bw, avx512vbmi and bmi2 flags. Oracle's fixed Java 17 build is 17.0.21.1. Several vendors also list Java 21, 25 and 27 as affected; others call those unconfirmed, so patch them too.
How do I fix Java4Shell? Install your vendor's fixed JDK (Oracle 17.0.21.1, 21.0.13.1, 25.0.5.1 or 27.0.1.1; Corretto 17.0.21.1.2 per Qualys; check Temurin, Zulu, Red Hat and your distro for their builds), restart every JVM and rebuild your Docker images. If you can't patch yet, set JAVA_TOOL_OPTIONS=-XX:-UseBASE64Intrinsics as a temporary, unvalidated stopgap.
Is Java4Shell being exploited? As of 17 to 18 November 2026, no exploitation in the wild had been reported and it was not in CISA's KEV catalogue. Researchers say they have built working exploits for several products and are holding them back, so assume public exploits will follow.
Are ARM servers and AWS Graviton affected by Java4Shell? They are reported as not affected, because the vulnerable code path uses x86 AVX-512 instructions. Still update them on your normal schedule, since Oracle's November update fixes other Java issues too. Do not push -XX:UseAVX=2 to ARM machines; OpenJDK on ARM reportedly fails to start with it.
Is Java4Shell related to Log4Shell? Only by name. Log4Shell (2021) was a flaw in the Log4j logging library. Java4Shell is reported to be in the Java runtime itself, so updating libraries in your pom.xml or build.gradle does not fix it; you have to update the JDK or JRE that runs your code.
If you're not sure how many JVMs are hiding in your stack, or you want a patch routine that makes the next one of these boring, book a 30-minute call. Bring your Dockerfiles and the list of tools your team self-hosts; that's usually where the Java turns up.
Sources
- Oracle Critical Security Patch Update Advisory, November 2026 (Oracle)
- CVE-2026-114938 record (CVE Program; reserved at the time of writing)
- How to Find, Mitigate, and Patch the Critical Java4Shell Flaw (CVE-2026-114938) Now (Qualys)
- Java4Shell: remote code execution in OpenJDK (CVE-2026-114938) (Datadog Security Labs)
- CVE-2026-114938: Reported Critical Pre-authentication Remote Code Execution in Java 17 (Rapid7)
- Java4Shell critical OpenJDK Base64 decoder flaw (Orca Security)
- Java4Shell (CVE-2026-114938): unauthenticated remote code execution in Java 17, 21, 25 and 27 (Cycode)
- Java4Shell Java RCE: Fix and Versions (CVE-2026-114938) (Aikido Security)
- Java4Shell (CVE-2026-114938): unauthenticated RCE reported in OpenJDK 17 (Endor Labs)
- Java4Shell: Oligo blocks the zero-day before the patch (Oligo Security)
- Java4Shell CVE-2026-114938: Java 17 RCE checks and hunting (ThreatDefence)
- Known Exploited Vulnerabilities Catalog (CISA)
Written By
Kunal Vohra
Technical Co-Founder & Fractional CTO
I've co-founded 6+ startups across India, the UAE, and the US, spanning AI, Web3, fintech, and cybersecurity. I write about the technical and strategic decisions that determine whether a startup thrives or stalls.
Comments
Loading comments…