Severity
Critical
CVSS
8.1 (CVSS v3.1), 9.2 (CVSS v4.0)
Status
Public PoC, no known exploitation
Fix
Patch available
Affected
Oracle Java SE, GraalVM for JDK, GraalVM Enterprise and OpenJDK 17, 21, 25 and 27 before the 17 November 2026 release, plus Oracle Java SE 8 performance builds, on x86-64 processors with AVX-512 VBMI

CVE-2026-114938, known as Java4Shell, is an out-of-bounds write in the fast path Java uses to decode Base64 data on recent Intel and AMD processors. A crafted Base64 string makes the decoder write past the memory it allocated, corrupting other data the application holds.

Calif, which reported the flaw, has shown it can be used to read application memory and to run code on the server. It can be exploited remotely without authentication.

Update to a fixed JDK build as soon as your vendor releases one. OpenJDK fixed the flaw in its 17 November 2026 release, and vendor builds and distribution packages are following. Until you can update, start affected JVMs with -XX:-UseBASE64Intrinsics, which removes the vulnerable code path. Systems on ARM64, or on x86-64 processors without AVX-512 VBMI, are not affected.

Exploitation status

We are not aware of any confirmed exploitation in the wild, and CVE-2026-114938 had not been added to CISA's Known Exploited Vulnerabilities Catalog as of 18 November 2026.

Calif's public proof of concept is a test case, not a weaponised exploit. It shows whether a Java version and processor hit the bug, not whether an application passes attacker input to the decoder. Calif has said exploits for specific products will follow later.

Kevin Breen of Immersive posted on X on 17 November that his team had built a working exploit and had to tailor it to each target application. He pointed to two routes that need no login: Apache Tomcat's Basic authentication handling and the java-jwt library. He has not said that the exploit will be released.

According to Calif, a failed attempt can crash the application, and an attacker can usually try again.

Root cause

Java's memory safety covers code written in Java, not the JVM itself. To speed up common operations, the HotSpot JVM swaps some standard library methods for hand-written assembly, known as intrinsics, when the processor supports faster instructions. On x86-64 processors with AVX-512 VBMI, java.util.Base64.Decoder uses one of these, and the bug is in the assembly that deals with the last few characters of the input.

For an input of 64 valid characters followed by a single =, Java allocates a 50-byte array for the result. After decoding the first 64 characters, the routine mishandles the trailing =, calculates a negative output length and performs a full 64-byte write, 62 bytes past the end of the array. Java rejects the input as invalid, but only after the write, by which point neighbouring objects on the heap may already be corrupted.

Calif describes two outcomes from the same corruption. JavaBleed, as Calif names it, leaks the contents of the process's memory, where an application keeps material such as signing keys, session tokens and credentials. The more serious outcome is full code execution in the context of the Java process.

Who is affected

All of the following must be true for an application to be exploitable:

  • Java version: it runs on an affected Java build. For OpenJDK, that means 17, 21, 25 or 27 before the 17 November 2026 release, and Red Hat puts the start of the affected range at 17.0.6. OpenJDK 8 and 11 are not affected.
  • Processor: the host is x86-64 and supports AVX-512 VBMI. This covers Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD Zen 4 and newer.
  • JVM flags: the JVM has not been started with -XX:-UseBASE64Intrinsics, or with -XX:UseAVX=2 or a lower AVX level.
  • Attacker-controlled input: the application passes Base64 data from an attacker to the standard library decoder. Common sources include HTTP Basic credentials, SAML artefacts, CSRF tokens, JSON Web Tokens and MIME encoded words, many of which are decoded before authentication completes.

Listening services aren't the only target. The researcher samczsun supplied Calif with a client-side exploit in which the victim is a Java program making an outbound TLS connection to a malicious server. That program doesn't need to accept any connections itself to be compromised, so features that call webhooks or fetch user-supplied URLs are in scope too.

Applications Calif tested

Calif achieved code execution against the products below, each running with its out-of-the-box settings on a vulnerable JDK and a processor with AVX-512 VBMI. Six of them could be attacked without credentials. Calif describes this as a first round of testing, so a product missing from the table hasn't been cleared.

ProductWhen exploitable
KeycloakBefore authentication
Atlassian JiraBefore authentication
Atlassian ConfluenceBefore authentication
JFrog ArtifactoryBefore authentication
JenkinsBefore authentication
PingFederateBefore authentication
Apache Kafka (broker)After authentication
Apache SolrAfter authentication
ElasticsearchAfter authentication
GerritAfter authentication
SpinnakerAfter authentication

Calif's testing went beyond packaged products to sample applications built on Apache Struts and Spring Boot, which were also exploitable. Login forms, file uploads and token handling all decode Base64, so in-house Java applications on an affected host should be treated as exposed until shown otherwise.

Affected and fixed versions

The OpenJDK Vulnerability Advisory of 17 November 2026 lists OpenJDK 17, 21, 25 and 27 as affected, up to and including the versions below. The advisory doesn't list fixed version numbers, so those below are taken from the fixed packages in Debian unstable.

OpenJDK branchAffected up toFixed version
OpenJDK 1717.0.21.0.317.0.21.1
OpenJDK 2121.0.13.0.321.0.13.1
OpenJDK 2525.0.5.0.325.0.5.1
OpenJDK 2727.0.1.0.327.0.1.1

Oracle lists the following supported versions of its own products as affected:

  • Oracle Java SE: 8u511b33-perf (performance build), 17.0.21.0.3, 21.0.13.0.3, 25.0.5.0.3 and 27.0.1.0.3
  • Oracle GraalVM for JDK: 17.0.21.0.3 and 21.0.13.0.3
  • Oracle GraalVM Enterprise Edition: 21.3.20.0.3

At the time of writing, Debian 12 (bookworm) and Debian 13 (trixie) are still listed as vulnerable. Red Hat lists Red Hat Enterprise Linux 8, 9 and 10 and the Red Hat build of OpenJDK 17, 21 and 25 as affected, with an update to follow. Other OpenJDK distributions release on their own schedules, so check your vendor's advisory for the fixed build.

How to check whether you are affected

Start with the processor, as a host without AVX-512 VBMI is not affected whatever Java version it runs. On Linux:

grep -o avx512vbmi /proc/cpuinfo | head -1

If the command prints avx512vbmi, the processor supports the vulnerable instruction set. If there is no output, it does not. According to Calif, the decoder takes the vulnerable path only when the processor reports all four of avx512f, avx512bw, avx512vbmi and bmi2.

On hosts that pass the processor check, check the Java version:

java -version

For a built image, check the runtime inside the image:

docker run --rm --entrypoint java your-image:tag -version

Mitigation

  • Apply the update: move to a fixed build from your JDK vendor as it becomes available, starting with internet-facing services and the applications listed above.
  • Rebuild containers: if an image is pinned to an older JDK base image tag, change the tag to a fixed release, then rebuild and redeploy.

Stop-gap mitigations

Until a fixed build is available, Red Hat recommends starting affected JVMs with one of these flags, which you should test before rolling out to production:

  • -XX:-UseBASE64Intrinsics: disables the hardware-accelerated Base64 decoder only. This is the preferred option, as it removes the vulnerable code path and leaves other AVX-512 optimisations in place.
  • -XX:UseAVX=2: disables AVX-512 instructions entirely. Apply it only to x86-64 hosts, as JVMs on ARM64 do not recognise it and will fail to start.

Either flag only addresses CVE-2026-114938. The same OpenJDK release fixes nine other CVEs across HotSpot, the security libraries and the core libraries, so plan the update even after applying a flag.

Update log

18 November 2026

Initial publication.

References