A futuristic cybersecurity illustration shows a protected server sending data to a glowing microchip.
Every so often a bug comes along that makes security teams rewrite their whole week. Java4Shell is that bug this week. CVE-2026-114938 is a memory-corruption flaw in OpenJDK's Base64 decoder. On the right CPU, it can crash a Java virtual machine, leak heap memory in the style of Heartbleed, or give an attacker remote code execution, in some cases before anyone logs in. Oracle shipped fixes in its November 2026 Critical Security Patch Update on Tuesday, November 17. According to The Register, CISA has urged Java users to patch right away.

If your Windows Server estate runs Jira, Confluence, Jenkins, Keycloak or any other Java software on modern Intel Xeon or AMD EPYC hardware, read on.

A futuristic cybersecurity illustration shows a protected server sending data to a glowing microchip. What Java4Shell actually is​

Base64 turns binary data into text. It shows up in HTTP Basic authentication headers, SAML artifacts, CSRF tokens, encoded email headers and file uploads. To decode it faster, OpenJDK's HotSpot JVM has a hand-written assembly routine called an intrinsic. It uses AVX-512 VBMI (Vector Byte Manipulation Instructions) on x86-64 processors that support them. The bug is in that routine.

Calif.io's write-up describes the mechanism:

  • The trigger: 64 "A" characters followed by one "=" sign. On OpenJDK 17 the trigger is 128 "A" characters plus "=", because that branch needs longer input before it reaches the vulnerable code.
  • The fault: For the 65-character input, Java allocates a 50-byte output array. The assembly routine mishandles the trailing padding character and calculates a negative output length. That turns its write mask into "write everything", so it writes a full 64 bytes and runs 62 bytes past the end of the array.
  • The timing problem: Java does reject the malformed input, but only after the neighbouring heap memory has been overwritten.
  • The JIT angle: The vulnerable intrinsic only runs once the C2 compiler has compiled the decoder as hot code. Calif's exploit "warms up" the decoder with repeated requests first.

Java4Shell is in the Java runtime itself, so no application or library has to be at fault for a service to be exposed. Log4Shell lived in a logging library you could find and swap out. This bug lives in the runtime underneath your applications, so you can't fix it by updating one dependency.

Who is affected: three conditions must all be true​

Red Hat's bulletin, RHSB-2026-018, lists the conditions for exploitation:

  1. The runtime: OpenJDK 17.0.6 or later, including the 21 and 25 lines. Red Hat says OpenJDK 8 and 11 don't contain the vulnerable code path.
  2. The CPU: x86-64 with AVX-512 VBMI. Red Hat names Intel Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, plus AMD Zen 4, Zen 5 and Zen 6. ARM64 systems and older x86 chips aren't known to be affected.
  3. The application: It passes Base64 data that an attacker controls to java.util.Base64.Decoder, and it hasn't been started with AVX-512 disabled.

Red Hat says no special configuration is needed: the fast path is chosen automatically on capable CPUs, so affected systems are vulnerable out of the box.

Red Hat also lists cloud instance families most likely to be affected. For Azure, these include Dsv5, Dsv6, Dasv6 and Dasv7 and related D-series v5–v7 SKUs. For AWS, C6i/M6i/R6i and C7i/M7i/R7a. For Google Cloud, C3, C3D, C4, C4D, N4, N4D, M3 and M4. Many Windows Server VMs in Azure run on exactly that hardware.

Upwind's lab testing supports the version boundaries. It reports affected: Temurin 17.0.6, 17.0.21, 21.0.13 and 25.0.5; clean: Java 11.0.33 and 17.0.5.

The Java 8 wrinkle​

Oracle's list of vulnerable builds includes 8u511b33-perf. Upwind explains this is its Performance Pack build, as affected. Standard Java 8 has no vulnerable decoder. If you run the Performance Pack edition of Java 8, it's in scope. Plain Java 8 isn't.

Which applications can be attacked​

Calif says it tested popular Java software in default configurations and sorts the results into two groups:

  • Pre-authentication RCE: Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate.
  • Post-authentication RCE: Apache Kafka Broker, Apache Solr, Elasticsearch, Gerrit and Spinnaker.

Calif says both lists are incomplete. It also demonstrated RCE in sample apps built on Apache Struts and Spring Boot, and it describes a client-side scenario in which a Java program connecting over TLS to a malicious server gets compromised.

Calif is also clear that there is no single exploit that works everywhere. Each target needs its own exploit chain. Its exploit was sensitive to network latency, so the team rented a VM in the same cloud region as the target to make the attack reliable. That doesn't make the bug harmless, but it does mean mass exploitation will take more work than Log4Shell did.

The impact is real, though. Before Calif reached code execution, it had what it calls "JavaBleed": the ability to read Java heap memory that may hold personal data, private keys and session tokens. Any code an attacker runs executes with the privileges of the Java process. That is one more reason not to run Jenkins as SYSTEM.

Severity: 9.2 or 8.1?​

The researchers who found it rate it 9.2 under CVSS v4.0, and Oracle rates it 8.1 under CVSS v3.1. These are scores from two different scoring systems, so they don't contradict each other. Oracle's risk matrix marks the flaw as remotely exploitable without authentication but with high attack complexity, which matches the exploit-chain details above.

Exploitation status​

As of November 17, Cisco Talos said it had not seen in-the-wild exploitation but expected probing to increase after disclosure. The Canadian Centre for Cyber Security said the same in alert AL26-029. ThreatDefence reported that no exploitation and no public exploit had been reported by November 18, 2026, and it is not in the CISA KEV catalog.

That may not last. Tenable notes that on November 17, Kevin Breen of Immersive said on X that they had developed a working exploit. Calif, for its part, is holding back its product-specific exploits for now. Cyber Kendra also found that GitHub repositories carrying the CVE ID have started to appear, although the one it reviewed only contained a version-checking script. Be careful with any "exploit" or "scanner" repository that appears this week. Fake tools around hot CVEs are a common way to deliver malware.

Calif also says details of the bug leaked before Oracle published its fixes. Upwind repeats that account. No independent source reviewed here has verified the leak, but it's a good reason not to wait for the next maintenance window.

The fix: patched builds and vendor status​

Oracle lists these vulnerable builds: Java SE 8u511b33-perf, 17.0.21.0.3, 21.0.13.0.3, 25.0.5.0.3 and 27.0.1.0.3; GraalVM for JDK 17.0.21.0.3 and 21.0.13.0.3; and GraalVM Enterprise Edition 21.3.20.0.3.

UC Berkeley's Information Security Office, citing Oracle and the OpenJDK advisory, lists these replacement builds:

Java lineUpgrade to
Java SE 8 / OpenJDK 88u513
Java SE 11 / OpenJDK 1111.0.33.1
Java SE 17 / OpenJDK 1717.0.21.1
Java SE 21 / OpenJDK 2121.0.13.1
Java SE 25 / OpenJDK 2525.0.5.1
Java SE 27 / OpenJDK 2727.0.1.1

Berkeley says Oracle advises moving every line, including 8 and 11, to these replacement releases. Most likely that's because the November update fixes 11 Java SE flaws, not just this one. For GraalVM and JavaFX, Berkeley's advice is to upgrade when patched versions become available.

Fix status at other Java vendors varies. The Hacker News reports that Amazon has fixed its Corretto Java builds, but Red Hat is still preparing fixes for its own. It adds that Azul's November update fixes the flaw in Zulu 17, 21, 25 and 27. Red Hat's bulletin was still marked "Ongoing" and lists RHEL 8, 9 and 10 plus its OpenJDK 17, 21 and 25 builds as affected.

Early reports also disagreed. Orca Security said Oracle hadn't yet confirmed the CVE was among its fixes. ThreatDefence said the CVE record was still reserved. Upwind says the flaw is now confirmed by OpenJDK and fixed in Oracle's 17 November builds. The risk matrix in Oracle's advisory does list CVE-2026-114938. If you use Temurin, Microsoft's Build of OpenJDK, Liberica or another distribution, check that vendor's own release notes rather than assuming a version number from Oracle covers you.

If you can't patch today: interim mitigations​

Each of these flags needs a JVM restart. Remove the flag once you've applied the real fix.

  • -XX:-UseBASE64Intrinsics is Red Hat's preferred option. It turns off only the hardware-accelerated Base64 code and keeps the other AVX-512 optimizations.
  • -XX:UseAVX=2 turns off AVX-512 entirely. The Canadian Cyber Centre and Talos both recommend it. Use it only on x86-64: ARM64 JVMs don't recognize it and will fail to start. Berkeley warns it can slow other workloads.
  • -XX:+UnlockDiagnosticVMOptions -XX:DisableIntrinsic=_base64_decodeBlock is an additional option Berkeley lists.

These are HotSpot JVM options, so they aren't specific to Linux. On Windows, add them wherever the JVM's startup options are set: a service wrapper configuration, a Tomcat or Jenkins service definition, or a vendor's launcher file. The location varies by product. Red Hat warns that for software that bundles its own JVM, you should apply the flag or update through the vendor's supported path, not by swapping the runtime yourself.

A practical triage plan for Windows and mixed estates​

  1. Inventory every Java runtime, including the vendor and full build number. Check bundled JREs inside commercial products, appliances and container images, not just system-wide installs. Running java -version against each runtime's own binary is the starting point. A server can have several JREs.
  2. Check the hardware. On Linux, Red Hat suggests grep -o avx512vbmi /proc/cpuinfo | head -1, and the Canadian Cyber Centre suggests lscpu | grep avx512. The sources here don't give a Windows equivalent. In practice, check the CPU generation against the Ice Lake / Zen 4-and-later list and your Azure VM SKU.
  3. Prioritize internet-facing identity and authentication services such as Keycloak, PingFederate, and the login pages of Jira and Confluence. Those services decode Base64 credentials before authentication completes.
  4. Patch, or apply a mitigation flag, then restart.
  5. Rebuild containers once updated base images are available.
  6. Watch your logs for repeated malformed Base64 in auth headers, unexpected JVM crashes, and Java processes spawning shells. Talos has published Snort rules (Snort2 SIDs 1:67502–1:67513; Snort3 additionally 1:301754) that detect malformed Base64 in relevant HTTP fields.

A Windows-specific caveat: a JVM crash on an Ice Lake server with no obvious cause should now get a security review, not just a quick restart.

The rest of Oracle's November update​

Java4Shell headlined a very large release. The Register counts 655 patches, 62 rated critical, across 62 products. Several score higher than Java4Shell:

  • Six scored a maximum 10.0: three in Oracle Access Manager, and one each in Oracle Platform Security for Java, WebLogic Server and Business Intelligence Enterprise Edition.
  • Five scored 9.9, in Oracle Text, Hyperion, the RDBMS and AutoVue.
  • Twenty-eight scored 9.8.
  • Tenable counts 249 high-severity fixes, and Hyperion alone accounts for 267 of the bugs.

WebLogic and Access Manager often sit on the internet edge in enterprise environments, so don't let Java4Shell push them down your list.

The awkward bits: hype, AI, and who pays for an optimization​

Not everyone liked how the bug was announced. Akamai's chief information security architect Jan Schaumann wrote on the Openwall mailing list that Calif had put "publicity-maxxing" first. He pointed to the opening line of its disclosure, a joke about people not knowing what PHP was after its PHP2Shell disclosure. The branding, website and New York Times coverage are a lot of marketing for a bug with a high-complexity attack. Still, the technical claims have been backed up by Red Hat, Cisco Talos, the Canadian government and independent lab tests from Upwind.

The more lasting point is about cost. Calif says Quang Luong found the bug in about an hour, built a first working exploit two hours later, and spent less than $200 on AI tokens. Calif also argues that the AVX-512 Base64 speed-up mainly helps applications that decode huge amounts of data, yet every Java deployment on capable hardware ships it and carries its risk. Expect more cases like this as AI tools make it cheap to audit hand-written assembly in language runtimes. Also expect vendors to be asked more often why a given optimization is switched on by default.

Bottom line​

Java4Shell is a serious flaw in the Java runtime. It only matters where three things line up: Java 17.0.6 or later, an AVX-512 VBMI processor, and an application that decodes Base64 from untrusted input. Plenty of modern servers meet all three. As of November 18, there was no confirmed exploitation in the wild, but one researcher already claims a working exploit and the disclosure reportedly leaked early.

Find your runtimes, patch through each vendor's supported path, use -XX:-UseBASE64Intrinsics where you can't patch yet, and check the rest of Oracle's November fixes while you're at it.

 

References

  1. RHSB-2026-018 Remote Code Execution via memory corruption on x86-64 systems - OpenJDK (CVE-2026-114938) - "java4shell" | Red Hat Customer Portal access.redhat.com
  2. CVE-2026-114938 OpenJDK and Oracle Java vulnerabilities | Information Security Office security.berkeley.edu
  3. Java4Shell CVE-2026-114938: Java 17 RCE Checks and Hunting threatdefence.com