November 18, 2026

Java4Shell: Which Java Servers Are Actually at Risk

At 1:00 p.m. Pacific time on Tuesday, November 17, 2026, Oracle and OpenJDK published fixes for CVE-2026-114938, a flaw in the Java runtime. Calif, the U.S. security firm that found it, named it “Java4Shell,” and the Cybersecurity and Infrastructure Security Agency (CISA) issued an alert urging immediate patching the same day. In Japan, IPA, JPCERT/CC and JVN each published notices on November 18, Japan time. Everything below reflects what was public as of the evening of November 18, Japan time (the morning of November 18 in New York).

What goes wrong

Java has a built-in Base64 decoder, the routine that turns text back into binary data. It sits behind login headers, file uploads and authentication tokens all over the web. To make it faster, Java’s HotSpot runtime has a special version that uses newer x86-64 CPU instructions (AVX-512 VBMI). The bug is in that special version.

According to Calif’s technical write-up, an input of just 65 characters, sixty-four “A”s followed by a single “=”, makes the decoder write 62 bytes past the end of a 50-byte buffer. Java does reject the input in the end, but by then the neighboring memory is already damaged. Calif says it turned that damage into two things: reading memory out remotely, in the same style as the 2014 Heartbleed bug in OpenSSL, and running arbitrary code. Memory can hold personal data, private keys and authentication tokens.

The weakness type is CWE-122, a heap-based buffer overflow (CVE record). Java is known for managing memory on the programmer’s behalf, which makes this class of bug rare. This time the flaw is not in the Java language but in hand-written machine code inside the runtime.

What Calif actually broke into

Calif lists Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate as products it took over without logging in, in their default configurations. With a login, it also succeeded against Apache Kafka, Apache Solr, Elasticsearch, Gerrit and Spinnaker. These are standard tools companies run on their own servers. It also succeeded against sample apps built on Apache Struts and Spring Boot, and notes that ordinary features such as login and file upload are the way in.

Calif also describes a client-side attack: a Java program that merely opens a TLS connection to a malicious server can be compromised. Java programs that never listen for connections are not automatically safe.

Calif says it found the bug by pointing Codex, OpenAI’s coding AI, at complex parts of OpenJDK. Finding it took about an hour, the first working exploit about two more, and the AI usage cost under $200. Calif is also the firm that disclosed PHP2Shell, the PHP flaw, in October. It is holding back its product-specific exploits until people have had time to update, but a short program that reproduces the memory corruption is already public.

Three conditions have to line up

The first condition is the Java version. In OpenJDK’s November 17 advisory, this CVE is marked against the 17, 21, 25 and 27 lines, not 8 or 11. Calif says the 17 line is affected from 17.0.6 on. The one exception is Oracle’s paid Java SE 8 Performance Edition (8u511b33-perf), which Oracle lists as affected.

The second is the CPU. The vulnerable code only runs on x86-64 processors with AVX-512 VBMI. Calif names Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD EPYC and Ryzen from Zen 4 on. ARM processors, such as AWS Graviton and Apple’s M chips, are not affected. In the cloud, Red Hat’s bulletin RHSB-2026-018 names the instance families most likely to qualify: AWS C6i, M6i, R6i, C7i, M7i and R7a; Google Cloud C3, C3D, C4, C4D, N4, N4D, M3 and M4; and Azure Dsv5, Dsv6, Dasv6 and Dasv7. In other words, the newer the server or cloud instance, the more likely it qualifies.

On Linux, Calif gives this check. If all four flags, avx512f, avx512bw, avx512vbmi and bmi2, come back, the CPU qualifies.

grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' | grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'

The third is that outside input reaches the decoder. Apps that decode Base64 before a user has logged in, at a login page, an API authentication step or a file upload, are the riskiest shape. Red Hat likewise says to start with internet-facing applications and authentication services.

Scored 8.1, but not by all

Oracle scores the flaw 8.1 under CVSS 3.1 (November update advisory). It is reachable over the network, without a login, all the way to takeover, but Oracle rates the attack complexity as high, which keeps it from the top of the scale. Others score it higher: Calif gives it 9.2 under the newer CVSS 4.0, Red Hat 9.8 and Atlassian 9.2. The difference comes down to how hard each company thinks the attack is.

No exploitation has been reported in the U.S. or Japan. In the CVE record, CISA marked exploitation as “none,” automation as “no” and technical impact as “total” (21:53 UTC on November 17). When I checked at 13:39 UTC on November 18, the flaw was not in CISA’s Known Exploited Vulnerabilities catalog. JPCERT/CC also says it had no information about real-world exploitation as of November 18.

Calif explains why automation is hard: each product needs its own exploit, down to how memory is laid out, and there is no one exploit that works everywhere. Still, the fix itself is public now. With PHP2Shell, exploitation was confirmed on October 29, eight days after PHP released its fix on October 21 (where PHP2Shell reports and the primary sources disagree). Whether you patch before exploitation starts is what matters this time.

How it differs from Log4Shell

The name invites comparison with Log4Shell (CVE-2021-44228) from December 2021. That was a flaw in the Log4j library, triggered just by writing a string to a log, regardless of CPU. It was easy to exploit and was attacked worldwide almost as soon as it was published.

Java4Shell is a flaw in the Java runtime itself, so any Java app that decodes Base64 can be in scope, whatever libraries it uses. In that sense it is broader than Log4Shell. But it comes with version and CPU conditions, and each product needs a tailored exploit. In that sense it is much narrower and harder. This is not a case where every company running Java is in immediate danger, as with Log4Shell. The right order is to find the servers that meet the conditions and fix those first.

Fixed versions

For Oracle’s paid Java SE, the fixed versions are 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. The free distributions built from OpenJDK are rolling out their own. Amazon Corretto shipped 17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2 and 27.0.1.1.4 on November 17 (the CVE appears in each version’s changelog, for example Corretto 21’s). Azul Zulu shipped the same day.

Others are behind. Eclipse Temurin (Adoptium) announced fixed builds 17.0.21.1+2, 21.0.13.1+2, 25.0.5.1+2 and 27.0.1.1+4, but when I checked at 13:40 UTC on November 18, only 27 was actually available. Container images come only after every version is out. For the Microsoft Build of OpenJDK, the “latest” download links now point to 17.0.21.1 and 25.0.5.1, but the link for 21 still pointed to the unfixed 21.0.13.0.2 when I checked at the same time. Red Hat lists RHEL 8, 9 and 10 and the Red Hat build of OpenJDK 17, 21 and 25 as affected and says fixes are being expedited. Ubuntu and Debian had not fixed their stable releases yet.

Even if you never installed Java yourself, a product may ship with its own. CISA warns that Java runtimes may not show up as separate entries in a software inventory. Here is where the products stand.

Atlassian says it will release out-of-cycle patches for the Data Center editions of Jira, Jira Service Management, Confluence, Bitbucket, Bamboo and Crowd once the Temurin builds they bundle are available (Atlassian community post). Until then it recommends the startup flag described below. It says Cloud customers need to take no action.

Keycloak’s official blog says container images 26.8.2-1 and 26.8.2-2 bundle an affected OpenJDK 21.0.13, and that the images will be rebuilt once OpenJDK fixes are out. Elasticsearch merged a change on November 18 that bumps its bundled Java to 27.0.1.1, but no release with it is out yet. For Jenkins, JFrog Artifactory and PingFederate, I found no notice about this flaw as of 13:40 UTC on November 18. Jenkins usually runs on a Java you installed yourself, in which case you update that Java.

In the cloud, Google Cloud’s bulletin GCP-2026-082, published November 17, says it has hidden the affected CPU instructions from the managed Java 17, 21 and 25 runtimes on App Engine standard and Cloud Run (second generation) through November 30. Other Cloud Run setups, Cloud Functions, Dataproc, Dataflow and customer-built containers need a setting on the customer’s side. AWS added Java4Shell detection to the AWS WAF Known Bad Inputs rule group on November 17, but the new rules only count matches by default; you have to change them to Block (AWS WAF changelog). For Azure Functions on Java, a Microsoft engineer explains that the default configuration does not use the vulnerable code.

Java 8 and 11 still need patching

IPA’s notice, Oracle Java vulnerability countermeasures (November 2026), lists Java SE 8 and 11 among the affected products. That looks like a contradiction with “8 and 11 are not affected,” but both are right. IPA’s list covers the whole November update, not just Java4Shell.

Oracle’s November update fixes 11 Java SE flaws, Java4Shell included, and Oracle says all 11 may be exploitable over a network without a login. Two of them, in TLS handling (JSSE) and Kerberos handling (JGSS), score 7.4 and also apply to 8 and 11. Being outside Java4Shell’s scope is no reason to skip this month’s update. Oracle now ships these updates monthly; the next one is December 15.

For companies in Japan

Many Japanese corporate systems have been built on Java for a long time, and plenty of core systems still run on Java 8 or 11. Those are outside Java4Shell’s scope. The systems to watch are the ones rebuilt or moved to the cloud in the last few years. Java 17 or later on a recent CPU generation is exactly the combination those migrations tend to produce.

There is also the question of Japanese middleware. Products such as Fujitsu Interstage, NEC WebOTX and Hitachi Cosminexus ship with their own Java runtime inside. As of the evening of November 18, Japan time, neither NEC’s nor Hitachi’s security notice index listed anything about this flaw. I could not load Fujitsu’s index, so I have not been able to check it. With products like these, you generally wait for the vendor’s fix rather than updating Java yourself. The fastest route is to ask the systems integrator or maintenance company three things: which product and which Java version are inside, what CPU it runs on, and when a fix is planned.

Japan’s public bodies treated this a notch lower than the U.S. did. JPCERT/CC published it as a CyberNewsFlash, not a formal alert. IPA rated its notice “caution” rather than “urgent.” JVN published JVNVU#99608534 on November 18; as of that day, its vendor status section listed a single company, Terumo BCT, as “no affected products (investigating).”

Japanese-language coverage contains a fair number of translations and summaries that differ from the originals. We checked the primary sources in the original language for this article, and we recommend that readers check them too.

If you cannot patch right away

Until you can install a fixed version, a Java startup flag switches the vulnerable code off. The one Calif gives, and Red Hat, Keycloak and JPCERT/CC also cite, is -XX:-UseBASE64Intrinsics. It disables only the fast Base64 path, and Red Hat and Keycloak recommend it. The other, -XX:UseAVX=2, disables AVX-512 altogether; Atlassian and Google Cloud recommend that one. Either way, Java has to be restarted for it to take effect.

One caution: Java on ARM does not recognize -XX:UseAVX=2 and will fail to start. In mixed ARM and x86-64 environments, add -XX:+IgnoreUnrecognizedVMOptions so unknown flags are ignored, as Atlassian and Google Cloud do.

CISA also lists validating Base64 input at the application boundary, limiting input size and limiting what can be reached without logging in. Simply moving internet-facing Jira, Confluence or Jenkins behind a VPN lowers the risk considerably. Calif notes that nginx’s default of buffering the full request body breaks some of its exploit steps, but says it does not stop the Elasticsearch chain and is not a mitigation.

What to do now

Start by listing every server running Java 17 or later. java -version shows the version, but Java bundled inside a product has to be checked where that product lives. Next, check those servers’ CPUs, with the command above or by cloud instance type. Among the servers that qualify, patch first the ones that face the internet and accept input before login, such as Jira, Confluence, Jenkins and Keycloak. Where the product’s fix is not out yet, apply the startup flag, then swap in the fixed version when it arrives and remove the flag.

After patching, keep watching for a while for Java crashing unexpectedly. CISA lists unexpected JVM crashes and unfamiliar Java runtimes or processes as signs of attack. When Java crashes, it leaves a file whose name starts with hs_err_pid. If an internet-facing server has been crashing for no clear reason, assume keys and tokens in memory may have been read and rotate them. How to respond when you suspect damage is covered in what to do in the first 24 hours after finding your site was hacked.

Who updates what, and how, is covered in keeping software up to date: learning about new versions, getting them and putting them live. We wrote about Jenkins core’s own flaws in October in 15 Jenkins core flaws. The background on AI-found vulnerabilities is in our recount of 2026 CVE registrations.

What is still unknown

The product-specific exploits have not been published, and Calif has not said when they will be. There are no dates yet for fixed releases from Atlassian, Keycloak or Elasticsearch. As of the evening of November 18, Japan time, I had not found anything from Japanese middleware vendors or Japanese cloud providers either.

Free security consultations

We are currently taking free consultations on security. Questions like "Is our site safe?" or "Where should we start checking?" are welcome.

Super Sonic Design accepts consultations on website creation, application design, etc. Please feel free to contact us.
inquiring
他の最新記事
カテゴリー