Content
Services
Industries
Talk to an expert

Java4Shell: the flaw is in Java, not in the application

· 7 min read · Network Secure

On 17 November 2026, OpenJDK and Oracle fixed CVE-2026-114938, nicknamed Java4Shell: an out-of-bounds memory write in the optimised Base64 decoder of the HotSpot JVM. It affects Java 17, 21, 25 and 27 on x86-64 processors with AVX-512 VBMI. Calif, which found and reported the flaw, reports unauthenticated remote code execution in products such as Keycloak, Jira and Jenkins in their default configuration.

The case is one day old, and sources still disagree on important points. So we separate what is confirmed by the advisories from what is still reported by researchers and vendors. This article is defensive: it does not include the input that triggers the flaw or any exploitation detail.

What Java4Shell is

Unlike Log4Shell in 2021, the flaw is not in a library but in the runtime itself. When code that decodes Base64 runs often enough, the JVM's JIT compiler replaces the Java code with an assembly routine that uses AVX-512 instructions. According to Calif, that routine miscalculates which bytes to write for the final part of a padded input (the = sign) and writes past the end of the output array, corrupting neighbouring objects in JVM memory.

Base64 is everywhere: authentication headers, tokens, SAML assertions, uploads and certificates. That is why Calif says ordinary features such as login and file upload can carry attacker data to the vulnerable decoder. Updating the application does not help. You have to update the Java that runs it.

What is confirmed and what is still reported

ClaimStatus on 18 Nov 2026
Fixed by OpenJDK and OracleConfirmed. Builds 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1, released on 17 November (CERT-EU advisory 2026-021).
SeverityDisputed. OpenJDK and Oracle assign CVSS 3.1 8.1 (High): network attack, no login, but high complexity (AV:N/AC:H/PR:N/UI:N). Calif calculates CVSS 4.0 9.2 (Critical).
US government alertConfirmed. On 17 November, CISA urged immediate patching, prioritising internet-facing systems.
Unauthenticated RCE in Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederateReported by Calif from initial testing in default configurations. In Kafka, Solr, Elasticsearch, Gerrit and Spinnaker, execution would require login. Calif says the list is not exhaustive.
Exploitation in real attacksNot confirmed. Sophos knew of no public reports on 18 November. Tenable cites attempts observed by GreyNoise, with no confirmed success.
Public proof of conceptPartly confirmed. Calif published a program that reproduces the memory corruption without taking control of the process. Application-specific exploits were held back, and Calif says it will publish them later.
CISA KEV catalogNot listed on 18 November 2026.

To see why the two scores differ, read how to read a CVSS score. To decide what to fix first when the score, KEV and EPSS point in different directions, read how to prioritise vulnerabilities.

Who is exposed

The flaw needs three conditions at the same time:

  • An affected Java. OpenJDK 17.0.6 or later on the 17 line, and the 21, 25 and 27 lines, in any distribution: Temurin, Corretto, Zulu, Microsoft, Liberica, SapMachine, Red Hat and Oracle. Oracle's Java SE 8 performance builds (8u511-perf), GraalVM for JDK 17 and 21 and GraalVM Enterprise Edition 21.3 are also affected. In the OpenJDK advisory, the CVE does not mark Java 8 or 11.
  • An x86-64 processor with AVX-512 VBMI. Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD Zen 4 or newer. On Linux, the CPU must report all four flags avx512f, avx512bw, avx512vbmi and bmi2. ARM64 is not affected.
  • Outside data reaching a Base64 decode. On a web server, API or identity provider, that is the norm. Calif also describes a client scenario: a Java program that opens a TLS connection to a malicious server.
LineAffected versionsFixed in
1717.0.6 to 17.0.2117.0.21.1
2121.0.13 and earlier21.0.13.1
2525.0.5 and earlier25.0.5.1
2727.0.1 and earlier27.0.1.1
Java SE 8, Oracle performance build8u511-perf8u513-perf

Distributions follow their own schedule. On 18 November, according to Tenable, Corretto and Zulu already had fixed builds, Temurin was starting to release its own, and Red Hat and Ubuntu had no packages yet; Debian had fixed only the unstable branch. Check your vendor's advisory, not just the upstream version number.

Java 8 and 11 also need updating, for another reason. The same 17 November release fixes nine more OpenJDK CVEs, and several affect all six lines, including 8 and 11, with scores up to 7.4. Being out of Java4Shell's reach does not mean there is nothing to patch.

Three inventory traps come up often:

  • The application's Java is not the system's. Container images, products with an embedded JRE (Jira, Jenkins, Keycloak, appliances) and build tools ship their own runtime.
  • The same image changes risk with the machine. The host CPU decides. In the cloud, changing the instance type changes the processor.
  • A running JVM keeps the old code until it restarts, even with the new package installed.

Interim mitigation

Updating is the fix. Until then, Calif and CERT-EU describe two JVM startup options that turn off the affected code. Neither Oracle nor OpenJDK has published them, and Sophos has not validated them. Treat them as a stopgap and test before using them in production.

  • -XX:+UnlockDiagnosticVMOptions -XX:-UseBASE64Intrinsics turns off only hardware-accelerated Base64 decoding. This is the form CERT-EU gives; Calif and Keycloak cite only -XX:-UseBASE64Intrinsics.
  • -XX:UseAVX=2 turns off AVX-512 entirely and may reduce performance. Use it only on x86-64: on ARM, the JVM will not start with this option.

Confirm the effect with java -XX:+PrintFlagsFinal -version, passing the same options as the service, and remove the option after updating.

Where even that is not possible, CISA recommends compensating controls: restrict and validate Base64 input at the application boundary, enforce size limits and reduce unauthenticated access.

What to hunt since 17 November

As with PHP2Shell, the patch closes the flaw but does not prove that nobody exploited it first. With no published IoCs, hunting starts from behaviour. CISA asks defenders to monitor for unexpected JVM crashes, rogue runtimes and suspicious Java processes. The hypotheses below detail those signals and are Network Secure's technical derivation, not published indicators.

  1. JVMs crashing for no apparent reason. hs_err_pid*.log files, segmentation faults and restarts of Java services on hosts with AVX-512 VBMI. A failed attempt to corrupt memory tends to bring the process down. Source: service logs, orchestrator restarts, EDR.
  2. A Java process starting a shell or utilities such as sh, bash, curl, wget or PowerShell (ATT&CK T1190, T1059). Source: EDR, auditd, Sysmon.
  3. New files in application, plugin or temporary directories, created by the service account after 17 November (T1505.003, T1105).
  4. New outbound traffic from Java servers to destinations with no history (T1071). Source: firewall, NDR, egress proxy.
  5. Malformed Base64 in authentication and upload fields, such as the Authorization header, tokens and SAML, with spikes of decoding errors grouped by source. Source: WAF, reverse proxy, application logs.
  6. Server credentials and tokens used elsewhere (T1552.001, T1078). Keycloak, Jenkins and Jira hold valuable secrets.

If there is any sign, treat it as an incident, not a patch. Preserve memory, disks and logs before restarting or updating: a restart erases exactly what the investigation needs.

What to do now

  1. Inventory every Java 17 or newer, including containers and products with an embedded JRE, with version, distribution, host CPU, exposure and owner.
  2. Prioritise internet-facing services with login, such as Keycloak and other identity providers, Jira, Confluence, Jenkins and Artifactory, running on hosts with AVX-512 VBMI.
  3. Update to your vendor's fixed build and rebuild images. Restart and confirm the version of the running process: java -version on the PATH may not be the service's runtime.
  4. Where you cannot update today, apply the JVM option and record the exception with a removal date.
  5. Track KEV, distribution advisories and those from product vendors. Publication of the application-specific exploits is likely to change the picture.
  6. Hunt from 17 November using the hypotheses above. To structure the patching process, see vulnerability management.

Frequently asked questions

Is Java4Shell the same problem as Log4Shell?

No. Log4Shell (CVE-2021-44228) was in the Log4j library. Java4Shell is in the HotSpot JVM itself, in the optimised Base64 decoder, and only exists on x86-64 processors with AVX-512 VBMI.

Are Java 8 or Java 11 affected?

Not by Java4Shell: in the OpenJDK advisory, the CVE affects only 17 from 17.0.6, 21, 25 and 27, plus Oracle's Java SE 8 performance builds (8u511-perf). But the same 17 November release fixes other CVEs that affect 8 and 11, so they should be updated too.

Are ARM servers exposed?

No. The flaw depends on AVX-512 instructions on x86-64 processors. ARM64 servers are not affected.

Does the -XX:-UseBASE64Intrinsics option replace the patch?

No. It is a stopgap described by Calif and CERT-EU, not published by Oracle or OpenJDK. The fix is to update to the fixed build and restart the JVM.

Sources consulted up to 18 November 2026: OpenJDK vulnerability advisory of 17 November; Oracle Critical Security Patch Update, November 2026; CVE-2026-114938 record on CVE.org; CISA alert of 17 November and KEV catalog; Calif's Java4Shell write-up; CERT-EU advisory 2026-021; Sophos SA-20261118-java4shell; Tenable FAQ (updated 18 November, with reports from GreyNoise, VulnCheck and Keycloak). Patch, KEV and exploitation status change quickly. The hunting hypotheses are Network Secure's technical derivation.

Official references

Read next