Java4Shell: Who the Java Flaw Affects and Who It Doesn't
In short
Java4Shell only hits Java 17.0.6 and later on x86-64 processors with AVX-512 VBMI, meaning Intel Xeon from Ice Lake and AMD from Zen 4. The flaw CVE-2026-114938 in the Base64 decoder has been fixed since 17 November 2026 in Java 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. Java 8 and 11 and ARM servers are not affected. Websites on PHP or Node.js are not affected either, but Java services next to them are, such as Elasticsearch, Jenkins, Keycloak or Jira. No attacks had been confirmed by 20 November.
Four weeks after PHP2Shell and Node2Shell, it is the third runtime’s turn: Java. The flaw is called Java4Shell, and the name brings back memories of Log4Shell in 2021. The comparison only holds in part. Here is what is behind it, who is affected and what to do. This article reflects 20 November 2026, and the sources are listed at the end.
What Java4Shell is
Java4Shell is the nickname for CVE-2026-114938, a bug in the Base64 decoder of OpenJDK. OpenJDK is the basis of practically every widely used Java runtime. Base64 is a way of transmitting arbitrary data as text. It is used in login headers, in login tokens and in many interfaces.
On certain processors, Java decodes Base64 with an accelerated, hand-written variant. According to the security firm Calif, it miscalculates when the input ends with the padding character = in a particular way. The decoder then writes 62 bytes past the end of a memory area. Java does report an error afterwards, but by then the write has already happened. This can crash the Java runtime, expose memory contents and, in the worst case, lead to the execution of foreign code.
The flaw was found by Quang Luong of Calif, the same firm that published PHP2Shell in October. Calif writes that the search ran with an AI tool and took about an hour. Oracle scores the flaw at 8.1 out of 10 (CVSS 3.1, “high”), Calif at 9.2 (CVSS 4.0, “critical”). Red Hat rates it as critical.
The timeline according to Calif:
- 9 June 2026: A user publicly reports a crash to the Java vendor Adoptium.
- 8 and 9 September: A developer finds the cause and publicly proposes a fix. The proposal is never merged and is later closed. The triggering input had therefore been online since September.
- 24 October: Calif finds the bug independently and reports it to Oracle, together with a working attack on Elasticsearch.
- 17 November: Oracle publishes the fixed versions. Calif publishes its analysis. The US agency CISA calls for immediate patching on the same day.
- 18 and 19 November: The other Java vendors follow, but not all at the same speed.
Three conditions have to come together
This is the difference from Log4Shell. Java4Shell only applies when all three points are true:
- Java 17.0.6 or later. The lines 17, 21, 25 and 27 are affected. The standard builds of Java 8 and 11 do not contain the faulty code. One exception is Oracle’s Java 8 Enterprise Performance Pack.
- An x86-64 processor with AVX-512 VBMI. According to Calif, that means Intel Xeon from the Ice Lake generation and AMD EPYC and Ryzen from Zen 4. ARM processors are not affected, so neither are the Graviton instances at AWS.
- Untrusted Base64 data reaches the decoder. This is the condition that is met most often. Many applications decode the login header, a token or a SAML message before they check whether anyone is logged in.
In the cloud there is one more catch: what matters is the processor the virtual machine passes through to Java. According to Palo Alto Networks and Keycloak, the same container image can therefore be affected on one node and not on the next.
Which programs are affected
Calif says it has demonstrated code execution without login against Keycloak, Atlassian Jira and Confluence, JFrog Artifactory, Jenkins and PingFederate, each in the default configuration. With a login, it succeeded against Apache Kafka, Apache Solr, Elasticsearch, Gerrit and Spinnaker. According to Calif, the list is not exhaustive.
Two caveats belong here. Calif has published a test that triggers the bug, but is holding back the finished attacks. And only part of this has been confirmed independently so far: Dynatrace writes that it has not verified the attack chains itself. Blackpoint Cyber was able to reproduce crashes and the reading of memory within the same process in its own tests, but not remote code execution.
The crash alone is already a problem. Blackpoint crashed a test application based on Spring Boot with around 28,000 malformed login requests in about 20 seconds, without credentials.
Is the flaw already being exploited?
As of 20 November 2026, no successful attack has been publicly confirmed. AWS writes on 19 November that it has no evidence of exploitation. Dynatrace reports scanning and exploitation attempts on the same day. According to Dynatrace, the flaw was not in the CISA catalogue of known exploited vulnerabilities on 19 November.
According to Calif, there is no attack that works everywhere. Each target needs its own preparation, and the attack is sensitive to network latency. Palo Alto Networks therefore does not expect broad, indiscriminate scanning like after Log4Shell. Upwind still expects others to build working attacks quickly.
What this means for typical website setups
WordPress, Laravel, static Astro website. No Java runs there. The website itself is not affected.
Search servers. Elasticsearch and Solr are Java programs and are on Calif’s list. If your shop or website search runs on one of them, look here first. For the managed OpenSearch Service, AWS writes that x86-64 domains on OpenSearch 2.7 or later may be affected. AWS applies the mitigation there itself.
Tools around the project. Jenkins for deployments, a self-hosted Jira or Confluence, Keycloak for login, Artifactory for packages. These services are often connected to the internet and have access to source code and credentials. They belong at the top of the list.
Hosted services. For cloud versions of Jira, Elasticsearch and the like, the provider is responsible. A look at its status page is still worthwhile.
AWS Lambda. According to AWS, the default environments do not support AVX-512 VBMI and are not affected.
What you should do now
- Find Java.
java -versionon servers and CI runners. For containers:docker run --rm --entrypoint java <image> -version. Many programs ship their own Java. According to Palo Alto Networks, they need an update from the vendor even if the system’s Java is already current. - Check the processor. On Linux:
grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' | grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'. If all four features appear, the processor is affected. - Update. The fixed versions are 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. Amazon Corretto is fixed from 17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2 and 27.0.1.1.4.
- Check the status of your vendor. Not every vendor was equally fast. According to Keycloak, the Linux builds of Eclipse Temurin were released on 19 November, with other platforms to follow. Red Hat wrote on 20 November that the packages for RHEL were being expedited. For Microsoft’s OpenJDK build and BellSoft Liberica, no fix was listed on 18 November, and according to Cycode, Debian only had the fix in its unstable branch at that point.
- Restart and rebuild. The old version keeps running until the Java process is restarted. Container images have to be rebuilt. According to Cycode, the Corretto images on Docker Hub keep their version tag, so pull again first.
- Do not rely on the application’s version number. Keycloak states explicitly that the Keycloak version alone says nothing about whether the Java inside it is fixed.
- Move off end-of-life Java lines. Java 18 to 20, 22 to 24 and 26 get no fix.
If you cannot update right away
The option -XX:-UseBASE64Intrinsics switches off the accelerated Base64 code. Red Hat names it as its preferred interim mitigation, AWS recommends it, and Upwind and Blackpoint have confirmed in tests that it prevents the bug. It can be set through the environment variable JAVA_TOOL_OPTIONS. Restart the process afterwards. Whether it has reached the running process is shown by jcmd <pid> VM.flags -all | grep -F UseBASE64Intrinsics.
Three notes on this:
- Sophos points out that neither Oracle nor OpenJDK itself has published a workaround.
- The option
-XX:DisableIntrinsic=_base64_decodeBlock, which is also circulating, prevents Java from starting, according to Upwind. The alternative-XX:UseAVX=2only works on x86-64. - A firewall rule does not replace the update. AWS writes about its own WAF rules that they only cover the flaw partially.
As warning signs, Palo Alto Networks and Upwind name unexplained crashes with files called hs_err_pid*.log, bursts of restarts and many error messages from the Base64 decoder (IllegalArgumentException). None of these proves an attack, and their absence does not prove the opposite.
What we do at Codeäffchen
Our website is a static Astro site, and no Java runs there. We build our projects with Laravel, WordPress and Astro. Java only appears there in services alongside, such as a search server. For those, the list above applies: find Java, check the processor, update, restart.
Three runtimes in four weeks. According to Calif, Java4Shell was found with an AI tool, and according to SecurityWeek, Node2Shell was reported by researchers at OpenAI. Finds like these are likely to become more frequent. The lesson stays the same: if you know where which runtime runs in which version, a day like this takes minutes. That also applies to the services next to the actual website.
Sources
- Calif: Java4Shell: Write once, RCE everywhere
- CISA: Critical Java Vulnerability Requires Immediate Patching
- Red Hat: Security Bulletin RHSB-2026-018
- AWS: Security Bulletin 2026-189-AWS
- Keycloak: Java4Shell: What Keycloak admins need to do
- Dynatrace: Java4Shell: Which Java runtimes are affected and how to fix them
- Blackpoint Cyber: Java4Shell: Java Base64 Decoder Flaw
- Palo Alto Networks: Java4Shell: Risks and Mitigations
- Upwind: Java4Shell: What We Know Now and What Comes Next
- Cycode: Java4Shell: unauthenticated remote code execution in Java 17, 21, 25 and 27
- Sophos: Advisory: OpenJDK Base64 Vulnerability
- The Hacker News: Java4Shell Flaw Enables Pre-Auth RCE in Jira, Jenkins, Keycloak and More
You run a web application with Java services in the background and are unsure whether it is affected? Find out more about our work on web applications on the software development page, or send us a short message.
Frequently Asked Questions
Is Java4Shell as bad as Log4Shell?
As things stand, no. In 2021, Log4Shell could be exploited with a single line of text against almost any Java application. Java4Shell needs a specific processor, Java 17.0.6 or later, and an attack built for each individual target. Palo Alto Networks therefore does not expect mass scanning like back then. The flaw still deserves attention, because the faulty code can be reachable before login.
Is Java4Shell already being exploited?
As of 20 November 2026, no successful attack has been publicly confirmed. AWS wrote on 19 November that it had no evidence of exploitation. Dynatrace reported scanning and exploitation attempts on the same day. According to Dynatrace, the flaw was not in the US agency CISA's catalogue of known exploited vulnerabilities on 19 November.
Which Java versions are safe?
Java 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1, released on 17 November 2026. Versions from 17.0.6 up to these releases are affected. Standard builds of Java 8 and 11 do not contain the faulty code. According to Cycode and Palo Alto Networks, the end-of-life lines 18 to 20, 22 to 24 and 26 do contain it and get no fix.
Is my WordPress, Laravel or Astro website affected?
The website itself is not, because no Java runs there. Services running next to it can be affected: a search server such as Elasticsearch or Solr, a Jenkins for deployments, Keycloak for login or a self-hosted Jira. The checklist in the article applies to those services.
How do I check whether my server is affected?
Check the Java version with java -version, including in containers and for programs that ship their own Java. On Linux, the file /proc/cpuinfo shows whether the processor reports the features avx512f, avx512bw, avx512vbmi and bmi2. The server is only affected if all four are present and Java 17.0.6 or later is running without the fix.
What can I do if there is no update for my Java yet?
Start Java with the option -XX:-UseBASE64Intrinsics. It only switches off the accelerated Base64 code. Red Hat names it as its preferred interim mitigation, and AWS recommends it as well. The Java process has to be restarted afterwards. The option does not replace the update.
Related Services
Founder & Full-Stack Developer
20+ years of web development experience. Specialised in Laravel, WordPress and custom software for mid-sized businesses.