Patch available CVSS 8.1 Security news
Java4Shell (CVE-2026-114938): Java 17+ Base64 decoder flaw allows unauthenticated code execution, its finder says
Oracle and the OpenJDK project on 17 November 2026 fixed CVE-2026-114938, which its finder, Calif, calls Java4Shell. The fast Base64 decoder in Java 17 and later can write past the end of a heap buffer on servers whose CPUs support AVX-512 VBMI. Calif says it turned the bug into unauthenticated code execution in Keycloak, Jira, Confluence, Jenkins and other products. Oracle and Amazon Corretto have shipped fixed builds. No in-the-wild exploitation has been reported so far.
Key points
- Who: server-side Java 17+ applications on x86-64 CPUs with AVX-512 VBMI. ARM and older CPUs are not affected.
- The bug: a malformed Base64 string makes the fast decoder write past the end of a heap buffer. The attacker only needs the application to decode their input, such as an Authorization header.
- Impact: Oracle rates it CVSS 3.1 8.1 (product takeover); Calif rates it CVSS 4.0 9.2 and reports unauthenticated code execution in Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate.
- Fix: Oracle and OpenJDK shipped fixes on 17 Nov 2026; Amazon Corretto's fixed builds are 17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2 and 27.0.1.1.4.
Vulnerabilities: CVE-2026-114938
What happened
At 1pm Pacific time (21:00 UTC) on Tuesday 17 November 2026, Oracle shipped fixes and the OpenJDK project published an advisory for CVE-2026-114938. The bug was found by Calif, which named it Java4Shell. It lives in the fast Base64 decoder inside the HotSpot virtual machine, so it is part of Java itself rather than a library. That means any application that decodes untrusted Base64 on an affected server is exposed.
In its NVD record Oracle places the flaw in the Compiler component and scores it CVSS 3.1 8.1 (AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H). An attacker can reach it over the network without logging in, but the attack is hard to pull off; if it succeeds, the attacker can take over the product. Calif scores it CVSS 4.0 9.2. The US agency CISA issued an alert the same day urging immediate patching, but as of 18 November it has not added the bug to its Known Exploited Vulnerabilities (KEV) catalog.
Technical details
- Where: HotSpot has an accelerated implementation (an intrinsic) of
java.util.Base64that uses AVX-512 VBMI instructions on x86-64. Calif says OpenJDK has used it since 17.0.6. - Bug class: memory corruption, specifically a write past the end of a heap output buffer. It is triggered by a crafted Base64 string. Calif says the input is very simple and fits into ordinary HTTP request fields.
- Conditions: an x86-64 CPU with AVX-512 VBMI (Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids, Granite Rapids; AMD EPYC and Ryzen Zen 4 or newer) and an application that decodes attacker-supplied Base64. Typical examples are an
Authorization: Basicheader, a JWT or API parameters, often before authentication. - Impact: Calif says it achieved unauthenticated code execution in default configurations of Keycloak, Atlassian Jira and Confluence, JFrog Artifactory, Jenkins and PingFederate. It lists authenticated code execution in Apache Kafka, Solr, Elasticsearch, Gerrit and Spinnaker. The published PoC only shows the memory overwrite. Calif is holding back the product-specific exploit chains, so we could not verify the impact claims ourselves. Early analyses by Upwind and Equixly saw a JVM crash as the more realistic outcome. VulnCheck reproduced the bug on an Ice Lake CPU and calls a crash trivial: a single request usually won't bring down a default server, but a burst of requests reliably did in its testing. Code execution is much harder, according to VulnCheck. It needs careful heap grooming plus an application-specific memory leak to defeat ASLR. VulnCheck says it has built a working RCE exploit but offers it only to paying customers.
Who is affected
- OpenJDK 17.0.6 and later and builds based on it: Oracle Java SE and OpenJDK, Amazon Corretto, Eclipse Temurin, Microsoft Build of OpenJDK, Azul Zulu, BellSoft Liberica, SAP SapMachine, Red Hat OpenJDK and Oracle GraalVM for JDK (list per Calif and oss-security).
- Oracle lists as affected Java SE 17.0.21.0.3, 21.0.13.0.3, 25.0.5.0.3, 27.0.1.0.3 and the 8u511b33-perf performance build, GraalVM for JDK 17.0.21.0.3 and 21.0.13.0.3, and GraalVM Enterprise Edition 21.3.20.0.3.
- x86-64 with AVX-512 VBMI only. ARM processors (Graviton, Ampere, Apple) are not affected, Calif says. On Linux, check the CPU flags with
grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' | grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'. A server is affected if the output lists all four. - Watch for hidden Java: applications with a bundled JRE, containers and managed cloud services.
How to tell if you were compromised
Neither the vendor nor Calif has published indicators of compromise, and nobody had documented in-the-wild exploitation as of 18 November. Cisco Talos has released Snort rules that look for malformed Base64 in HTTP request fields: Snort 2 SIDs 1:67502-1:67513 and Snort 3 SIDs 1:301754 and 1:67502-1:67513. We also suggest watching for:
- unexpected JVM crashes and new
hs_err_pid*.logfiles, especially on internet-facing services, - unusually long Base64 values in
Authorizationheaders and tokens in proxy or WAF logs, made up of a single repeated character, - shell processes (
sh,bash,cmd.exe,powershell) started under the Java application server's account.
You can pull indicators out of your own data with our IOC extractor.
What to do
- Update Java. On 17 November Amazon Corretto shipped fixed builds 17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2 and 27.0.1.1.4 (based on OpenJDK 27.0.1.1+4). Oracle shipped its fixes the same day in its Security Patch Update. For other distributions (Temurin, Zulu, Red Hat, Microsoft), look for releases dated 17 November or later.
- Restart every JVM and check the version in the running process:
jcmd <PID> VM.version. Don't forget container base images and runtimes bundled with third-party products. - Until you can patch: turn off the fast decoder with
-XX:+UnlockDiagnosticVMOptions -XX:-UseBASE64Intrinsics(it is a diagnostic flag, so the JVM won't start without the first option), or set-XX:UseAVX=2on x86-64 only (a JVM on ARM will refuse to start with that flag). Talos also recommends disabling AVX-512. Both can slow down Base64 and cryptography, so test them first. - Prioritise internet-facing services that decode Base64 before login: SSO (Keycloak, PingFederate), Jira, Confluence, Jenkins and Artifactory.
- Track fixes from vendors that ship their own Java (Atlassian, JFrog, CloudBees) and beware of fake "exploits" on GitHub.
Timeline
- 17 Nov 2026: Upwind and Equixly published the first analyses of the reported bug.
- 17 Nov 2026, 21:00 UTC: Oracle shipped its Security Patch Update, OpenJDK published its advisory and Calif its write-up.
- 17 Nov 2026: Amazon Corretto shipped fixed builds, Cisco Talos released Snort rules and CISA issued an alert. The CVE-2026-114938 record appeared in NVD.
- 17 Nov 2026: VulnCheck confirmed an easy JVM crash and announced a private RCE exploit.
- 18 Nov 2026: not in KEV; no exploitation reported.
Update
18 Nov 2026, 10:00 UTC - addendum. We added VulnCheck's analysis: a burst of requests can easily crash the service, code execution needs heap grooming and a memory leak, and VulnCheck holds a working exploit it has not released. We also corrected the interim mitigation: -XX:-UseBASE64Intrinsics is a diagnostic flag and only works together with -XX:+UnlockDiagnosticVMOptions. No in-the-wild exploitation has been reported.
18 Nov 2026, 07:30 UTC - correction, now verified. The first version of this story on 17 November relied on the Upwind and Equixly analyses. It said nobody had shown code execution and that a JVM crash was the realistic impact. That no longer holds. The finder, Calif, describes unauthenticated code execution in Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate in its advisory, though it has not released the exploits. We verified the fixed builds from Amazon Corretto's changelogs and the affected builds from Oracle's NVD record. We added a more precise CPU check (four flags instead of one), the -XX:UseAVX=2 flag and the Talos detection rules. The story is now marked as verified.
17 Nov 2026, 22:50 UTC - NVD published Oracle's record for CVE-2026-114938, with a CVSS 3.1 score of 8.1 and a list of affected Java SE and GraalVM builds.
FAQ
I run Java on ARM (AWS Graviton, Apple Silicon). Am I affected?
According to Calif, no. The bug sits in a fast path that HotSpot only uses on x86-64 with AVX-512 VBMI. Patch anyway: the same update fixes other Java flaws too.
Is updating the system Java enough?
Not always. Many products bundle their own Java runtime or run in containers built on old base images. Check the version inside the running process (the command jcmd PID VM.version) and restart the JVM after updating.
Why did an earlier version of this story say the main risk was a JVM crash?
The first analyses by Upwind and Equixly on 17 Nov did not demonstrate code execution. The finder, Calif, then published its own advisory describing code execution in specific products, while keeping the exploit chains private. We corrected the story on 18 Nov (see Update).
Sources
- Calif - Java4Shell (CVE-2026-114938), finder's advisory (17 Nov 2026)
- NVD - CVE-2026-114938 (Oracle record and CVSS 3.1 score)
- Oracle - Security Patch Update, November 2026
- Amazon Corretto 21 - changelog (21.0.13.1.2, 17 Nov 2026)
- Cisco Talos - Java4Shell detection and guidance (17 Nov 2026)
- VulnCheck - Java4Shell CVE-2026-114938: Fine, I Guess We'll Do This Again (17 Nov 2026)
- oss-security - CVE-2026-114938 announcement (17 Nov 2026)
- CISA - Critical Java Vulnerability Requires Immediate Patching (17 Nov 2026)
- Upwind - Java4Shell: what we know now and what comes next (17 Nov 2026)
- Equixly - analysis of the Base64 decoder bug in Java 17, 21 and 25 (17 Nov 2026)
Spotted a mistake? Let us know and we will correct it openly.