Back to Blog
Insights

Java4Shell (CVE-2026-114938): Pre-Auth RCE in the OpenJDK Base64 Decoder

Java4Shell (CVE-2026-114938): Pre-Auth RCE in the OpenJDK Base64 Decoder

On 17 November 2026, Oracle shipped its November Java security release and the OpenJDK Vulnerability Group published an out-of-cycle advisory. The headline issue is CVE-2026-114938, an out-of-bounds write in the OpenJDK Base64 decoder that researchers demonstrated as unauthenticated remote code execution in default configurations of stock enterprise applications. Calif, the firm that discovered it, named it Java4Shell. Wiz Research independently built a working exploit within two hours of the details surfacing. If you run Java in production, this is an upgrade-and-restart event, and the sooner the better.

To be clear about provenance: IKZERO did not discover this vulnerability. This is our technical read of the Calif disclosure, the OpenJDK advisory, the Oracle CVE record and the Wiz analysis, written for teams who ship Java to production.

The bug in one paragraph

OpenJDK accelerates Base64 decoding on x86-64 processors that support AVX-512 VBMI using hand-written assembly, introduced by JDK-8268276 and backported to the Java 17 line via JDK-8296735, shipping in 17.0.6 in January 2023. The assembly mishandles the padding tail of a Base64 string. On inputs where the final chunk is only padding or padding adjacent characters (=, ==, X=, X==, XY=, ===), the output length calculation underflows to a negative value, a subsequent mask computation passes that negative value through unchanged, and a masked vector write (VMOVDQU8) stores all 64 bytes of a vector register instead of a handful. With the canonical 65 byte trigger, 64 A characters followed by =, that is 62 bytes written past the end of a 50 byte heap buffer. Java rejects the malformed input afterwards, but the memory is already corrupted. On OpenJDK 17 the block threshold differs, so the trigger there is 128 A characters followed by =.

The code had prior form. JDK-8273108 recorded segfaults in unrelated JVM machinery after the intrinsic landed, and JDK-8321599 documented data loss in the same decoding path. This is the third known defect in a routine that ships to every Java deployment on capable hardware.

From heap corruption to code execution

A fixed value overflow where only the landing position is controllable sounds hard to exploit. The disclosed chain turns it into full RCE in four stages:

  1. The writer. Through heap grooming, the overflow is aimed at the length field of an adjacent array. Once that array believes it is larger than it is, the attacker can keep writing chosen bytes into it, upgrading a fixed value corruption into a controlled linear write.
  2. The reader. A second groomed array has its length corrupted the same way, and its contents are copied back into an application response. That yields a heap disclosure primitive Calif calls JavaBleed, in the style of Heartbleed. Java heap memory holds session tokens, credentials and keys, so this stage is serious on its own.
  3. Repair. Disclosed original bytes are used to restore damaged object headers so the JVM survives, allowing retries and data-only attacks against application logic.
  4. Escalation. The controlled write is finally aimed at JVM internals, reaching the code cache and runtime hooks for code execution.

Remote grooming works by spraying object pairs across G1 garbage collector regions and pacing partial HTTP request bodies with TCP acknowledgements to control allocation order across the network. Failed attempts usually crash the JVM, which is noisy but retryable.

What the researchers demonstrated

Pre-authentication RCE in default configurations:

  • Keycloak
  • Atlassian Jira
  • Atlassian Confluence
  • JFrog Artifactory
  • Jenkins
  • PingFederate

Post-authentication RCE: Apache Kafka, Apache Solr, Elasticsearch, Gerrit and Netflix Spinnaker. Sample applications built on Spring Boot and Apache Struts were also exploited. samczsun additionally demonstrated a client side scenario: a Java program that merely opens a TLS connection to a malicious server can be compromised, even when it exposes no service of its own.

Wiz Research confirmed the story independently. Their AI research harness produced a working exploit within two hours of the details emerging and proved remote execution against Jenkins, Confluence and Apache Tomcat. Their most important operational finding: there is no universal exploit. Every target needs a bespoke chain matched to its Base64 usage, heap layout and response behaviour. That raises the bar for mass exploitation, but it does not protect packaged applications such as Jenkins or Jira, where builds and memory layouts repeat across thousands of deployments.

Are you affected?

Software: OpenJDK 17.0.6 and later, on every distribution built from it: Eclipse Temurin, Amazon Corretto, Azul Zulu, BellSoft Liberica, SAP SapMachine, the Red Hat build, the Microsoft build and Oracle Java SE. The upstream advisory lists lines 17, 21, 25 and 27 as affected; Java 8 and 11 never received this intrinsic, although Oracle separately lists its Java SE 8u511b33 Enterprise Performance build as affected. Oracle builds named in the CVE record: 17.0.21.0.3, 21.0.13.0.3, 25.0.5.0.3, 27.0.1.0.3, plus GraalVM equivalents.

Hardware: the vulnerable path only executes on x86-64 with AVX-512 VBMI. That means Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD EPYC or Ryzen from Zen 4 onward. ARM64 is not affected, which includes Graviton, Ampere and Apple Silicon.

Check the hardware gate on Linux:

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

All four flags must be present for the host to run the vulnerable code.

Patch and mitigate

Apply your vendor's fixed build immediately and restart every JVM process. Replacing files does not update running processes, and container base images must be rebuilt rather than patched in place.

LineFixed build (Oracle)
Java 88u513
Java 1111.0.33.1
Java 1717.0.21.1
Java 2121.0.13.1
Java 2525.0.5.1
Java 2727.0.1.1

Temurin, Corretto, Zulu and the rest ship on their own schedules; watch their advisories for the matching fixed build numbers.

If patching will lag on internet-facing systems, either flag removes the vulnerable code path:

-XX:-UseBASE64Intrinsics     # recommended, disables the accelerated decoder
-XX:UseAVX=2                 # x86-64 only; do not set on ARM hosts

Two partial measures, not fixes: nginx request-body buffering defeats some grooming profiles, but not all of them, and a WAF can catch specific payloads but cannot cover every path that feeds attacker input into a Base64 decoder. Restrict external access to Java management portals while updates roll out.

The scoring, briefly

Oracle's official CVSS 3.1 base score is 8.1 (AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H), with successful attacks resulting in takeover of Oracle Java SE. Calif scores the same bug 9.2 under CVSS 4.0. Both are correct; the high attack complexity in 3.1 reflects the per-target heap grooming, not any requirement for credentials or user interaction. CISA's enrichment classifies it as CWE-122, heap based buffer overflow, with technical impact rated total. As of publication there were no confirmed exploitation reports in the wild and the issue was not in the KEV catalog, but with patch and disclosure on the same day, assume that window will close quickly.

Detecting exploitation attempts

Failed grooming usually crashes the JVM. Sweep your estate for fresh hs_err_pid*.log files, SIGSEGV and SIGABRT exits on Java processes, and container exit codes 134 and 139. Unexplained crash patterns on unpatched, internet-facing Java hosts deserve investigation as possible trigger attempts. And treat unknown "exploit" repositories for this CVE as hostile; malicious fake PoC repositories are the standard follow-up to a disclosure like this.

Credits

Discovery, exploitation research and the technical disclosure are the work of Quang Luong at Calif, with exploit improvements by Jenish Rakholiya and artwork by Michal Zalewski. samczsun demonstrated the client side variant. Wiz Research independently confirmed remote exploitation. Oracle produced the patches and published the CVE record as assigning CNA; the OpenJDK Vulnerability Group released the out-of-cycle advisory and formally acknowledged Quang Luong. This article is IKZERO's independent analysis of their work.

Our team runs offensive testing and detection engineering for organisations that ship Java to production. If you need help validating exposure across your fleet, see our penetration testing, DevSecOps and SOC as a service offerings, or contact us directly.

Frequently Asked Questions

Is Java4Shell exploitable without authentication?

Yes, on affected builds and hardware. Researchers demonstrated pre-auth RCE in default configurations of Keycloak, Jenkins, Jira, Confluence, Artifactory and PingFederate. There is no universal point-and-shoot exploit, though: each target needs a bespoke chain, and only a harmless proof of concept trigger has been published so far.

Which Java versions are affected?

OpenJDK 17.0.6 and later, across the 17, 21, 25 and 27 lines and every downstream distribution, plus Oracle's Java SE 8u511b33 Enterprise Performance build. The vulnerable code only executes on x86-64 processors with AVX-512 VBMI, so ARM64 systems are not affected.

What is the official CVSS score for CVE-2026-114938?

Oracle's official score is 8.1 under CVSS 3.1 with high attack complexity. Calif scores it 9.2 under CVSS 4.0. The difference is the scoring system, not a disagreement about severity: both describe takeover-class impact from an unauthenticated network attacker.

Is CVE-2026-114938 being exploited in the wild?

No exploitation has been confirmed as of publication. The issue is not in the CISA KEV catalog and CISA's SSVC assessment records exploitation as none. Given the demonstrated chains and the short patch-to-weaponization window seen with PHP2Shell, patch immediately rather than waiting for that to change.

What is the fastest mitigation if we cannot patch today?

Start the JVM with -XX:-UseBASE64Intrinsics, which disables the vulnerable accelerated decoder, and restrict external access to internet-facing Java management portals. The flag is a stopgap with a performance cost on heavy Base64 workloads; the patch is the fix.

Get Started

Want a security-first build?

Get a free security review. We'll look at where you stand today and tell you what to fix first, no strings attached.

Talk to an Expert
Java4Shell (CVE-2026-114938): Pre-Auth RCE in the OpenJDK Base64 Decoder - IKZERO