Java4Shell flaw puts Java servers at risk of code execution
Java4Shell is a Java vulnerability that can let attackers without an account run their own code on widely used applications, including Jira, Jenkins and Keycloak. Security research firm Calif disclosed the flaw, tracked as CVE-2026-114938, and rates it 9.2 out of 10 under CVSS 4.0, a system for scoring software vulnerabilities.
The vulnerable code is present in Java 17.0.6 and later on the Java 17 line, as well as Java 21, 25 and 27. It runs only on x86-64 Intel and AMD processors that support AVX-512 VBMI, a processor feature used to speed up certain calculations. Affected chips include Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids processors, plus AMD Zen 4 and newer processors. ARM64 systems do not run this code, so the same container image, a packaged application and its dependencies, may be vulnerable on one server but not another.
The flaw is in Java's faster Base64 decoder. Base64 is a way of representing data using text, and servers commonly decode it before checking whether someone is logged in. Examples include HTTP Basic login headers, SAML artifacts and CSRF tokens. A malformed string ending with an extra equals sign causes the decoder to write 62 or 63 bytes beyond the end of its output. Java rejects the input, but only after the unwanted write has happened.
The data written past the boundary is fixed, which initially suggested that the flaw would mainly cause crashes. Calif says attackers can instead arrange the server's memory so that the overflow expands a neighbouring array. They can then use one array to write data of their choosing and another to read memory back, a technique the researchers call JavaBleed. Calif says this can reach Java components that execute code. There is no universal exploit: every target needs its own memory layout and timing, and some attacks are affected by network delays.
Calif says the attacks worked without a login, under default settings, against Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate. With a login, the researchers tested Apache Kafka Broker, Apache Solr, Elasticsearch, Gerrit and Spinnaker. They describe those lists as incomplete. A proxy that waits for complete requests, as nginx does by default, blocks some attacks but not all, and a failed attempt can crash Java before an attacker tries again.
The researchers have not released attack code for the six applications. Other security teams have reproduced crashes and a memory leak, but not code execution. Upwind reproduced memory corruption on an AWS server with an Intel Ice Lake processor, where the overflow damaged the next object in memory and eventually crashed Java. Blackpoint Cyber's Adversary Pursuit Group crashed a test Spring Boot application with about 28,300 malformed login requests over roughly 20 seconds, without logging in. In a local test, the group used JavaBleed to read a secret stored beside the overflow in all 30 runs, but did not reproduce remote code execution.
The commercial Java vendor scores the vulnerability 8.1 out of 10 under CVSS 3.1 and describes it as difficult to exploit but requiring no login. Its CVE record says successful attacks can result in takeover of its Java SE. Red Hat gives it a preliminary 9.8, while Amazon rates it 8.8 and describes the risk as a Java Virtual Machine crash for unauthenticated users or, in some application environments, code execution by an authenticated low-privileged user. Canada's Cyber Centre also describes remote exploitation without authentication. The flaw was not listed in CISA's Known Exploited Vulnerabilities catalog when checked, and the Cyber Centre said it was not aware of attacks exploiting it.
The commercial Java vendor and the OpenJDK project released fixes this week. The commercial Java vendor's fixed builds are Java 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. The Java 8 Performance Pack is fixed in 8u513. GraalVM fixes are GraalVM for JDK 17.0.21.1, GraalVM for JDK 21.0.13.1 and GraalVM Enterprise Edition 21.3.20.1. Standard Java 8 and Java 11 are not affected, but Java SE 8 Performance Pack build 8u511b33-perf is.
Amazon fixed Corretto 17, 21, 25 and 27 in builds 17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2 and 27.0.1.1.4. Azul fixed Zulu 17, 21, 25 and 27 in free builds 17.70.201, 21.54.201, 25.38.201 and 27.30.201. Red Hat had not released fixes for Red Hat Enterprise Linux 8, 9 or 10, or its OpenJDK 17, 21 and 25 builds. SAP released SapMachine 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1, but its release pages did not specify what they fixed. Eclipse Temurin, the Microsoft Build of OpenJDK and BellSoft Liberica had not listed fixed builds when checked.
Administrators should update bundled Java runtimes through the product maker, then restart Java services. Container images based on updated runtimes must be rebuilt and redeployed so running processes use the fixed code. Until an update is possible, the Java Virtual Machine can be started with -XX:-UseBASE64Intrinsics, which disables only the faster Base64 path, or with -XX:UseAVX=2, which disables AVX-512 entirely on x86-64 servers. The second option is not suitable for ARM64 because Java does not recognise it there and will fail to start. Upwind said both settings stopped the overflow in every affected version it tested, while warning that -XX:DisableIntrinsic=_base64_decodeBlock alone prevents Java from starting.
Upwind recommends looking for Base64 tokens of at least 32 characters that end in an equals sign but whose length is not a multiple of four, especially in Authorization headers. Other warning signs include bursts of IllegalArgumentException errors from java.util.Base64$Decoder, unexplained hs_err_pid crash logs and repeated restarts. These signs do not prove a compromise, and their absence does not prove a server is safe. Calif also says a Java program that opens a TLS connection to a malicious server can be compromised, although the researchers have not published how that path reaches the decoder. The exploit for that case came from researcher samczsun.
A malformed input was posted publicly on September 9 in a proposed OpenJDK pull request describing an overflow in the same AVX-512 Base64 decoder. The proposal was closed without being merged in early November after an automated notice said it had been inactive for more than eight weeks. For organisations running affected Java applications, AEU-I provides security-first IT, infrastructure and consulting that can support broader security planning around this kind of server vulnerability.