CIRCL Publications and Presentations

TR-104 - Critical Memory Corruption in OpenJDK Base64 Decoder (Java4Shell, CVE-2026-114938)

Introduction

On 17 November 2026, OpenJDK and Oracle published security updates addressing CVE-2026-114938, commonly referred to as Java4Shell. The vulnerability is a heap-based out-of-bounds write in the x86-64 hardware-accelerated Base64 decoder of the HotSpot Java Virtual Machine (JVM). Specially crafted Base64 input can corrupt Java heap memory and, under exploitable application conditions, enable information disclosure, denial of service or remote code execution (RCE) without authentication.

The vulnerability was reported by Calif and affects a shared Java runtime component rather than a single web application. Calif reports successful pre-authentication RCE demonstrations against several widely deployed Java products. Oracle rates CVE-2026-114938 8.1 (High) under CVSS v3.1, while Calif rates it 9.2 (Critical) under CVSS v4.0. CISA has called for immediate patching.

CIRCL recommends urgent identification and remediation of exposed Java runtimes, prioritising Internet-facing applications, identity infrastructure and services processing untrusted Base64 data. Exposure must be assessed against the actual JVM build, CPU capabilities and reachable application functionality; not all Java installations are affected.

Technical details

The issue affects the java.util.Base64.Decoder execution path accelerated by HotSpot’s x86-64 AVX-512 VBMI intrinsic. In the affected routine, malformed Base64 padding near the end of an input can cause an output-length calculation to underflow. The resulting invalid write mask permits an out-of-bounds write beyond the destination array. The Java decoder may subsequently reject the invalid input, but heap corruption has already occurred.

According to Calif’s technical analysis, a malformed 65-character input (64 A characters followed by =) can produce a 62-byte overwrite beyond the intended output buffer on an affected build and processor. Some older vulnerable Java 17 builds use a longer trigger. Exploitation requires the affected optimised code path to be selected, typically after the JIT compiler has compiled the Base64 decoder. Exploitability therefore depends on CPU support, JVM configuration and how the application processes attacker-controlled input.

A successful application-specific exploit may allow an attacker to disclose heap contents, corrupt adjacent objects, crash the JVM or execute code with the privileges of the affected Java process. The published research includes a bounded local reproducer and descriptions of application-specific exploitation techniques; it does not establish one universal remote exploit for every Java application.

Affected systems and vulnerability reachability

Affected runtimes

The vulnerable optimised decoder is present in OpenJDK 17.0.6 and later unpatched builds of the relevant Java release lines, including Java 17, 21, 25 and 27. OpenJDK-derived distributions may be affected, including Oracle Java, Amazon Corretto, Eclipse Temurin, Microsoft Build of OpenJDK, Azul Zulu, BellSoft Liberica, SAP SapMachine, Red Hat OpenJDK and applicable GraalVM distributions. The presence and availability of fixes are vendor- and build-specific.

Java runtime / release line CVE-2026-114938 status or remediation target
OpenJDK / Oracle Java 17 Affected from 17.0.6 in unpatched builds; fixed upstream in 17.0.21.1.
OpenJDK / Oracle Java 21 Unpatched builds affected; fixed upstream in 21.0.13.1.
OpenJDK / Oracle Java 25 Unpatched builds affected; fixed upstream in 25.0.5.1.
OpenJDK / Oracle Java 27 Unpatched builds affected; fixed upstream in 27.0.1.1.
Oracle Java 8 Enterprise Performance Pack Oracle specifically lists 8u511b33-perf as affected; the corresponding fixed release is 8u513-perf.
Standard OpenJDK 8 / 11 Not affected by this specific vulnerable intrinsic according to Red Hat; other November 2026 Java security fixes still apply.
Oracle GraalVM for JDK / GraalVM Enterprise Edition Oracle lists affected builds; follow Oracle’s specific supported fix or vendor mitigation guidance.

The fixed upstream version numbers above are not a substitute for distribution-specific advisories. A vendor may backport the correction without adopting the identical upstream version string. Conversely, updating an application’s libraries without replacing its bundled vulnerable JVM does not remediate this flaw.

Conditions for exploitation

All of the following conditions are relevant to reachability:

  1. Vulnerable build: the service executes on a JVM containing the unpatched Base64 intrinsic.
  2. Eligible processor: the runtime executes on x86-64 with AVX-512 VBMI, with the required instruction support exposed to the JVM. Calif’s Linux check expects avx512f, avx512bw, avx512vbmi and bmi2. Typical affected processor families include Intel Xeon Ice Lake and later server families, and AMD Zen 4 or later.
  3. Vulnerable intrinsic enabled: the HotSpot Base64 intrinsic and the necessary AVX-512 path have not been disabled.
  4. Attacker-controlled data reaches the decoder: an application or library decodes untrusted Base64 through the affected runtime path. Depending on the application, this may occur before authentication, after authentication, or when a client processes data returned by an untrusted service.
Deployment scenario Vulnerability reachable?
Affected unpatched JVM, x86-64 with AVX-512 VBMI, untrusted Base64 decoded through the intrinsic Yes, potentially; exploitability depends on the application.
Same CPU and JVM, but no attacker-controlled input reaches the affected decoder No identified remote path through this vulnerability; verify data flows.
Affected JVM on ARM64 No, for the reported x86-64 intrinsic.
Affected JVM on x86-64 without AVX-512 VBMI exposed No, for the reported AVX-512 VBMI path.
Affected JVM started with -XX:-UseBASE64Intrinsics Vulnerable code path disabled, if the flag is supported and effective.
Affected x86-64 JVM started with -XX:UseAVX=2 Vulnerable AVX-512 path disabled, if the flag is effective.
Vendor-confirmed, corrected JVM build No, for CVE-2026-114938.

Calif reports pre-authentication RCE demonstrations in default configurations of Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate. It also reports post-authentication exploit chains against other products. These demonstrations are evidence of affected application paths, not proof that every installation of each named product is exploitable. A reverse proxy, Web Application Firewall (WAF) or request buffering must not be considered a general fix: exposure can remain via other endpoints, application protocols or complete requests.

Verification

First identify the actual Java runtime used by each process. A system’s default java executable may differ from the JVM bundled inside an application, appliance or container.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
# Default Java installation (initial inventory only)
java -version
readlink -f "$(command -v java)"

# Running Java processes and their command lines (Linux)
ps -eo pid,args | grep '[j]ava'

# Where a JDK's diagnostic tools are available and permitted
jcmd -l
jcmd <PID> VM.command_line
jcmd <PID> VM.flags -all | grep -E 'UseBASE64Intrinsics|UseAVX'

Check processor support where the JVM executes, accounting for virtual machines and container scheduling:

1
2
3
# All four flags should be present for the reported vulnerable CPU path.
grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' | \
  grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'

Python script - Inventory visible JVMs for CIRCL TR-104 / CVE-2026-114938

Review how untrusted data reaches Base64 decoding, including authentication headers, credentials, tokens, uploads, SAML messages and application-specific APIs. Identify both externally exposed services and Java clients communicating with untrusted peers. Match runtime builds to vendor advisories and installed package revisions, not merely the Java major version. Do not treat the presence of AVX-512 flags alone as proof of exploitability, or their absence on one host as proof that an autoscaled deployment is safe.

Avoid running publicly available exploit or crash-triggering tests against production services as an inventory method. The researcher’s bounded local reproducer does not establish that a remote application is exploitable.

Mitigation and recommendations

1. Apply security updates immediately

  • Update affected Java runtimes to corrected, supported builds. For the upstream OpenJDK lines, the relevant fixed versions include 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1; consult the distribution’s security advisory for the corresponding patched package.
  • Prioritise Internet-facing authentication systems, CI/CD services, enterprise application servers and Java services that consume untrusted Base64 input.
  • Update vendor-bundled runtimes through the product vendor’s supported process. Rebuild container images and redeploy workloads; ensure all Java processes have restarted onto the corrected runtime.
  • Review Java 8, Java 11 and other Java installations for separate security fixes in the same update cycle, even if they are not affected by this specific Base64 flaw.

2. Disable the vulnerable intrinsic when patching is delayed

As a temporary risk-reduction measure, start affected HotSpot JVMs with one of the following flags:

1
-XX:-UseBASE64Intrinsics

This selectively disables hardware-accelerated Base64 decoding and is generally preferable to disabling broader CPU optimisations. Alternatively, on x86-64 only:

1
-XX:UseAVX=2

This disables AVX-512 use more broadly and may have wider performance implications. The latter option may be unsupported on ARM64 JVMs and can prevent startup. Validate vendor support, add the chosen option to the actual service/container JVM startup configuration, restart the Java process, and verify its effective flags. Neither option removes the underlying vulnerable code; deploy the security update as the permanent remediation.

3. Reduce reachable attack surfaces

  • Where feasible, restrict public access to unpatched Java services and administrative interfaces.
  • Validate and constrain Base64 inputs at application boundaries, enforce size limits and reduce unauthenticated functionality. These are defence-in-depth controls, not comprehensive mitigations.
  • Review application-specific vendor recommendations. Do not assume that request buffering, a WAF rule, TLS termination or a patched Java library prevents exploitation of an affected JVM.

4. Monitor and investigate suspicious activity

Look for unusual JVM crashes or restarts; abnormal decoding or authentication errors; unexpectedly spawned child processes; new scheduled tasks or web-shell artefacts; anomalous outbound connections; and unexplained changes to application files or secrets. Correlate application, reverse-proxy, authentication, host and endpoint logs where available.

These are investigative leads, not validated indicators of compromise for CVE-2026-114938. A successful exploit may leave little application-level evidence. Preserve forensic evidence and investigate potential credential exposure or persistence before restoring affected systems. Patching does not undo a prior compromise.

Exploitation status

As of 18 November 2026, the Calif disclosure includes a publicly available, bounded local proof of concept and reports successful application-specific pre-authentication and post-authentication RCE demonstrations. The researchers stated that they were withholding weaponised product-specific exploit chains at publication. These demonstrations do not by themselves establish attacks against production victims. The Canadian Centre for Cyber Security reported no known active exploitation in its 17 November 2026 alert. CIRCL has not independently confirmed widespread exploitation in the wild from the sources reviewed for this report.

Local KEV monitoring: CIRCL will update its local Known Exploited Vulnerabilities (KEV) records in Vulnerability-Lookup if credible evidence of exploitation of CVE-2026-114938 is observed within CIRCL’s constituency or the Luxembourg area. These local KEV records reflect observed exploitation evidence and may be updated independently of international KEV catalogues.

Guidance for organisations in Luxembourg

CIRCL strongly recommends that organisations identify and patch affected Java deployments without delay, particularly services used for authentication, software development and deployment, enterprise applications and externally reachable APIs. Validate the runtime, CPU and application-input path together rather than assuming either that all Java systems are vulnerable or that Java services behind proxies are protected.

Where immediate updating is not possible, disable the vulnerable intrinsic and apply network and input-validation restrictions while coordinating an urgent supported update. Organisations observing suspicious activity should preserve evidence, activate their incident-response procedures and contact their relevant CSIRT, including CIRCL where applicable.

References

Classification of this document

TLP:CLEAR — Information may be distributed without restriction, subject to copyright controls.

Revision

  • Version 1.1 – TLP:CLEAR – Added direct CVE hyperlink and local KEV monitoring commitment – 18 November 2026.
  • Version 1.0 – TLP:CLEAR – Initial version – 18 November 2026.
Top