Insight

Java4Shell (CVE-2026-114938): The OpenJDK Base64 RCE

Java4Shell is a pre-auth heap overflow in OpenJDK’s AVX-512 Base64 decoder. Here is the confirmed mechanism, which builds are fixed, and what to check on your Java hosts.

Java4Shell (CVE-2026-114938): The OpenJDK Base64 RCE

Changelog

  • 18 Nov 2026, 10:49 NZDT: Oracle published its full advisory and the CVE record, OpenJDK published its own advisory, and Calif published the technical write-up. The bug is confirmed. We rewrote the post around the confirmed mechanism, replaced the patch table with the builds that have actually shipped, corrected the version history we had speculated about, added Calif’s trigger string and the second mitigation flag, and ran Calif’s proof of concept on our own hosts.
  • 18 Nov 2026, 08:00 NZDT: Published.

Java4Shell is CVE-2026-114938, a heap buffer overflow in the Base64 decoder built into OpenJDK. A remote attacker with no credentials can corrupt heap memory, read it back in the style of Heartbleed, and reach code execution. It leaked before the fix, and as of today it is confirmed: Oracle, OpenJDK and the finder, Calif, have all published.

If you run Java 17 or newer on x86-64 servers with AVX-512, treat any that decode attacker-reachable Base64 as exploitable, and patch or mitigate now.

What is confirmed

The CVE record is published. Oracle, as the assigning authority, places the flaw in the Java compiler component and scores it CVSS 3.1 8.1 (vector AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H). The “high” attack complexity is why the score sits at 8.1 rather than higher: landing the overwrite takes work, and a miss can crash the JVM.

Calif’s write-up gives the mechanism. OpenJDK decodes Base64 in plain Java, but on x86-64 processors with AVX-512 the HotSpot C2 compiler swaps in a hand-written assembly routine (an intrinsic) once the decoder runs hot. That routine miscalculates the write mask for a malformed final block. For an input such as 64 A characters followed by a single =, it computes an output length of minus one, which leaves the 64-bit write mask fully open. The decoder then writes 64 bytes into a 50-byte buffer, 62 bytes past the end, before the Java layer rejects the input as invalid. By then the neighbouring heap objects are already corrupted.

Calif demonstrated pre-authentication remote code execution, in default configurations, against Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate, and post-authentication execution against Apache Kafka, Apache Solr, Elasticsearch, Gerrit and Spinnaker. A client-side variant can compromise a Java program that merely opens a TLS connection to a malicious server. The common thread is any path where outside Base64 reaches the decoder, which covers HTTP Basic auth, JSON Web Tokens, SAML and file uploads. Calif found the bug with an AI coding agent in about an hour and built the first exploit in two more, at a token cost it puts under $200. The New York Times covered the disclosure.

Calif is holding back the weaponised per-product chains until people have patched. Upwind independently reproduced the 62-byte overflow. At the time of writing there is no public exploit and no reported exploitation in the wild, the NVD entry is not yet populated, and the CVE is not in CISA’s Known Exploited Vulnerabilities catalog.

Which versions

Calif states Java4Shell affects OpenJDK 17.0.6 and later. That is the release where the AVX-512 decoder reached Java 17. (An earlier draft of this post speculated about 17.0.7 and about an older decoder bug, JDK-8321599; both were wrong. The “fixed in 17.0.11” claim circulating elsewhere refers to an unrelated data-loss bug in the same decoder, which has nothing to do with Java4Shell.)

The OpenJDK advisory lists CVE-2026-114938 against the hotspot/compiler component and marks Java 17, 21, 25 and 27 as affected. Standard Java 8 and 11 do not carry the AVX-512 decoder and are not marked. Oracle’s one exception is the Java 8 Enterprise Performance Pack (8u511b33-perf), which does include it and is listed as affected.

Because the bug lives in OpenJDK source, every distribution built from it is affected on the right hardware: Oracle Java SE, Amazon Corretto, Eclipse Temurin, Microsoft Build of OpenJDK, Azul Zulu, BellSoft Liberica, SAP SapMachine and Red Hat’s OpenJDK among them.

The hardware gate is real. The vulnerable code only runs on an x86-64 CPU with AVX-512 VBMI, which in practice means Intel Xeon from Ice Lake onward (Sapphire Rapids, Emerald Rapids, Granite Rapids) and AMD EPYC or Ryzen on Zen 4 or newer. Arm cores, including AWS Graviton, never run it.

Patch status

Status checked: 18 November 2026, 10:49 NZDT (17 November, 21:49 UTC). Oracle released the fixed builds, and Amazon Corretto shipped the same day. The other OpenJDK vendors have not published fixed builds or advisories yet.

DistributionFixed Java 17 buildStatus
Oracle JDK17.0.21.1Released, with 8u513, 11.0.33.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1 for the other lines
Amazon Corretto17.0.21.1.2Released 17 Nov. 21.0.13.1.2 and 25.0.5.1.2 released the same day
Eclipse Temurinnot yetLatest is 17.0.21+9 from 22 Oct. No fix published
Azul Zulunot yetLatest is 17.0.21.0.3. No fix published
Microsoft Build of OpenJDKnot yetLatest is 17.0.21.0.2. No fix published
Red Hatnot yetNo CVE page live yet
Debiannot yetNo tracker entry live yet
Ubuntunot yetNo page live yet
SUSEnot yetNo page live yet
Amazon Linuxnot yetNo page live yet

A fixed Oracle build does nothing for a Temurin or Zulu runtime, and a patched host does nothing for a JDK baked into a container image. Inventory by the runtime each service actually loads, down into your container images.

Check your hosts

Start with which Java is running, and where. A JDK bundled inside an application or a container image will not show up in your package manager.

# every running JVM on this host, with its main class
jcmd -l

# exact build of one of them
jcmd <pid> VM.version

# inside a container
docker exec my-container java -version

Then check the CPU. Calif’s test wants all four AVX-512 feature flags present:

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

If any of the four is missing, that host cannot run the vulnerable path. The surest local test is Calif’s proof of concept, a self-contained trigger that keeps every write inside a real array, so it is safe to run. It warms the decoder, feeds it the malformed input, and prints REPRODUCED on a vulnerable build. We ran it on OpenJDK 17.0.21, 21.0.13 and 27.0.1 on a host with AVX2 but no AVX-512, and all three printed NOT REPRODUCED with zero bytes changed, which is the hardware gate doing its job. On AVX-512 hardware an unpatched build will corrupt the 62 bytes.

Finally, list what the network can reach, and match the process IDs against the jcmd list:

sudo ss -lntp | grep java

Mitigation if you cannot patch yet

For distributions without a fixed build, Calif gives two JVM switches. Either one takes the decoder off the vulnerable routine:

# force the plain Java Base64 decoder
java -XX:-UseBASE64Intrinsics -jar app.jar

# or turn AVX-512 off entirely (x86-64 only; OpenJDK on Arm rejects this flag)
java -XX:UseAVX=2 -jar app.jar

Both are ordinary product flags that apply straight from the command line, despite early write-ups that called for a diagnostic switch first. We confirmed both start cleanly on OpenJDK 17.0.21. For a service you cannot easily edit, JAVA_TOOL_OPTIONS applies the flag at the next start:

[Service]
Environment=JAVA_TOOL_OPTIONS=-XX:-UseBASE64Intrinsics

The JVM prints Picked up JAVA_TOOL_OPTIONS on startup, and jcmd <pid> VM.flags then shows the flag is set. Both switches need a restart and will slow Base64-heavy workloads, so test first. A mitigated host is holding until its fixed build lands; it is not done.

The ordinary controls still apply on top. Take Java services that do not need to be public off the internet, and stop exposing version-revealing endpoints such as Spring Boot Actuator. Calif notes that a reverse proxy which buffers the full request body, as nginx does by default, breaks some of their grooming paths, but it is not a general mitigation: single-request targets such as Elasticsearch are unaffected by it.

Signs of a failed attempt

A miss usually crashes the JVM. Look for hs_err_pid<pid>.log files in the service’s working directory, and for containers that exit with code 134 or 139. A Java service that starts crashing this week with no deploy behind it deserves a look at the request that came just before.

What is still open

The remaining gaps are the other vendors’ fixed builds (Temurin, Zulu, Microsoft, Red Hat and the Linux distributions), the NVD enrichment, and the weaponised exploit chains Calif is holding until people upgrade. Both the Swiss and South Korean national cyber centres have told operators to inventory Java and patch. Calif’s last disclosure of this kind, a pair of PHP bugs in October, drew exploitation within a week of leaking. The gap between a public mechanism and a working exploit is short, so patch the hosts you can and mitigate the rest today.