Java4Shell: Critical OpenJDK Base64 Flaw Patched as Researchers Demo Remote Code Execution

Calif's Java4Shell flaw, CVE-2026-114938, lets malformed Base64 input corrupt OpenJDK heap memory on AVX-512 servers. Oracle and Amazon shipped fixed Java builds on November 17.
Official OpenJDK logo on a white panel above a dark terminal window showing a CPU feature check and the Java4Shell proof-of-concept result, with a neon magenta glow

A critical memory-corruption flaw in OpenJDK, the open-source code behind most Java runtimes, can let a remote attacker corrupt heap memory with nothing more than a malformed Base64 string. Security firm Calif, which discovered and reported the bug, calls it Java4Shell. It is tracked as CVE-2026-114938, and patched builds started shipping on November 17, 2026.

Calif said on X at 11:11 a.m. Eastern on November 17 that the vulnerability had leaked, and that it would publish technical details after Oracle released a patch. Its full write-up went live shortly after 4 p.m. Eastern. Calif scores the flaw 9.2 out of 10 (Critical) under CVSS 4.0 and says it demonstrated remote code execution in several widely used Java products in their default configurations.

What is confirmed so far

The OpenJDK Vulnerability Group lists CVE-2026-114938 first in its November 17, 2026 advisory. The advisory places the bug in the HotSpot compiler component and rates it 8.1 under CVSS 3.1, with a vector showing network access, no privileges, no user interaction and high impact to confidentiality, integrity and availability. The “high” attack complexity in that vector is the main reason the official score is lower than Calif’s figure.

According to the advisory, the flaw affects the Java 17, 21, 25 and 27 release lines. Java 8 and Java 11 are not marked as affected in the OpenJDK risk matrix. The advisory names the affected builds as 27.0.1.0.3, 25.0.5.0.3, 21.0.13.0.3, 17.0.21.0.3 and earlier, and its acknowledgements credit Quang Luong, the Calif researcher behind the write-up.

Oracle’s November 2026 Critical Security Patch Update includes 11 new security patches for Oracle Java SE. Oracle says all 11 can be exploited over a network without credentials, and the highest CVSS 3.1 score among them is 8.1. That matches the top score in the OpenJDK advisory, which lists ten OpenJDK CVEs and one OpenJFX CVE. When BreachNews checked shortly after 4 p.m. Eastern, Oracle’s page still showed the pre-release announcement, which does not name individual CVEs. Calif’s write-up links to Oracle’s JDK 17.0.21.1 release notes as the patch and thanks Oracle for sharing affected version details before publication.

How the Base64 bug works

Base64 is a common way to turn binary data into plain text. Java applications use it for login headers, tokens, file uploads and many kinds of network messages. To speed up decoding, OpenJDK’s HotSpot virtual machine swaps in a hand-tuned routine, known as an intrinsic, that uses AVX-512 VBMI instructions on supported x86-64 processors.

Calif says the bug sits in the part of that routine that handles the last few characters of the input. When the leftover input is malformed padding, such as a single stray =, the routine calculates an output length of minus one or minus two. That invalid value produces a write mask that is fully open, so the processor writes all 64 bytes of a vector register to memory instead of the few bytes it should write.

In Calif’s example, 64 letter A characters followed by one = make Java allocate a 50-byte result array. The decoder then writes 62 bytes past the end of that array. Java does reject the input in the end and throws an error, but by then the neighboring memory may already be corrupted. On older Java 17 builds the trigger needs 128 A characters before the =, because the fast path only starts at a longer input length.

The decoder also has to be compiled by the C2 just-in-time compiler first. In practice, that means the vulnerable code path is used once an application has decoded enough Base64 data for HotSpot to optimize it. Busy servers are the most likely to be in that state.

From a fixed write to code execution

On its own, the overflow is limited. Calif says an attacker controls where the stray bytes land, but not most of their values. The firm says it got around that by placing the decoder’s output next to another array and overwriting that array’s length field. That turned the fixed write into a larger write with chosen values.

From there, Calif describes building a second primitive that copies heap contents back into the application’s response. The firm calls this step “JavaBleed” and compares it to Heartbleed, because it can leak data from the Java heap, which may hold personal data, private keys, secret keys and session tokens. Calif says it then expanded the write further to reach the JVM’s code cache and gain code execution.

The attack is not a single point-and-shoot payload. Calif says each target needs its own profile, including a route to the decoder, a way to read data back and a matching heap layout. It also says the technique is sensitive to network timing, and that a bad placement can crash the JVM, although an attacker can often retry.

Products Calif says it compromised

Calif says its initial testing showed remote code execution, in default configurations, in these products:

  • Before authentication: Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate.
  • After authentication: Apache Kafka Broker, Apache Solr, Elasticsearch, Gerrit and Netflix Spinnaker.

The firm says the lists are not complete. It also reports demonstrating code execution in sample applications built on Apache Struts and Spring Boot, and says a client-side exploit shared by researcher samczsun can compromise a Java program that simply opens a TLS connection to a malicious server. These are Calif’s lab results. None of the named vendors had published their own advisory for this flaw when BreachNews reviewed the write-up.

Proof of concept and exploitation status

Calif published a short, self-contained Java proof of concept. It warms up the decoder, feeds it the malformed input and checks whether bytes past the decoded output were changed. Calif says every write in that test lands inside a larger array, so it is safe to run. The firm says it is holding back its product-specific exploit chains until users have had time to upgrade.

No in-the-wild exploitation has been publicly reported. CVE-2026-114938 was not in the CISA Known Exploited Vulnerabilities catalog as of the November 17 catalog update, and its CVE record had not yet been published on CVE.org or the National Vulnerability Database when BreachNews checked. Calif has not explained how details of the bug leaked before the planned disclosure.

The public trigger lowers the bar for anyone trying to build a working exploit. Calif’s earlier PHP2Shell disclosure was later followed by attacks on Check Point gateways.

Who is affected

Calif says Java4Shell affects OpenJDK 17.0.6 and newer. Distributions built from OpenJDK inherit the bug. Calif names Amazon Corretto, Eclipse Temurin, Microsoft Build of OpenJDK, Azul Zulu, BellSoft Liberica, SAP SapMachine, Red Hat OpenJDK, Oracle OpenJDK, Oracle Java SE and Oracle GraalVM for JDK.

Hardware matters just as much as version. The vulnerable routine only runs on x86-64 processors with AVX-512 VBMI support. Calif lists Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, along with AMD EPYC and Ryzen chips from Zen 4 onward. ARM64 processors are not affected. A server also has to pass attacker-controlled Base64 data into Java’s standard decoder, which Calif says ordinary features such as authentication and file uploads can do.

On Linux, Calif suggests this command to check whether a CPU has the required features. All four flags must appear for the host to be exposed:

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

Run the check where the JVM actually runs. In a virtual machine, the CPU features exposed to the guest are the ones that count.

Patches and mitigations

Oracle’s fixed Java SE builds are 27.0.1.1, 25.0.5.1, 21.0.13.1, 17.0.21.1, 11.0.33.1 and 8u513, all dated November 17, 2026. Amazon also released Corretto 17.0.21.1.2, 21.0.13.1.2 and 25.0.5.1.2 on November 17, and the Corretto change logs list CVE-2026-114938 among the fixes. Other OpenJDK vendors release on their own schedules, so teams should confirm the fixed version with each vendor they use, including Java runtimes bundled inside third-party products and container images.

Where an upgrade cannot happen right away, Calif says either of these JVM startup options blocks the bug:

  • -XX:-UseBASE64Intrinsics turns off the hardware-accelerated Base64 routines.
  • -XX:UseAVX=2 turns off AVX-512. Calif warns that this flag should only be used on x86-64 hosts, because OpenJDK on ARM does not recognize it and will fail to start.

Both options need a restart and may slow down Base64-heavy workloads. They are stopgaps, not a replacement for the patched build. Teams should start with internet-facing identity, build and collaboration servers, such as single sign-on, CI/CD and ticketing systems, since those are the kinds of products Calif says it compromised before login.

What to watch next

The key open questions are when the official CVE record is published, how quickly the other OpenJDK vendors and Linux distributions ship fixed packages, and whether vendors such as Atlassian, Jenkins, Keycloak and JFrog issue their own guidance. Calif’s decision to release product-specific exploits later also sets a clock for defenders. BreachNews will update this story if exploitation is reported or if CISA adds the flaw to its catalog.

Picture of m00s3c

m00s3c

Moose (@m00s3c) is the author of BreachNews, focusing on data breach intelligence, dark web monitoring, and threat analysis. His work involves analyzing breach claims, reviewing leaked datasets, and tracking threat actor activity to provide clear, factual reporting.

Latest News

BREACHNEWS.COM/SUPPORT/

Support Independent News.

Help support breach monitoring, investigations, infrastructure, and reporting.

Support the site →