OpenJDK fixes Base64 flaw CVE-2026-114938: which JVMs run on x86 hosts with AVX-512
OpenJDK published a vulnerability advisory on 17 November 2026 fixing CVE-2026-114938, a flaw in HotSpot’s Base64 decoding that the reporting team at Calif named Java4Shell. It affects unpatched JDK 17.0.6 and later; standard OpenJDK 8 and 11 do not contain the code. It only matters when the JVM runs on an x86-64 CPU with AVX-512 VBMI, and ARM64 is not affected. The score is 8.1 with the vector CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H, so no authentication is needed. CISA put out an alert the same day urging immediate patching.
Java4Shell invites comparison with Log4Shell, but the affected set is much smaller, and the hosts inside it are easy to overlook. The same container image is fine on Graviton and in scope on an Ice Lake or Zen 4 x86 instance, so neither the image nor the JDK version gives you the answer. The researchers say they demonstrated unauthenticated remote code execution against the default configurations of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate, reached through ordinary features such as login and file upload. As of 08:25 Taipei time on 18 November I found no report of exploitation in the wild and KEV had not added the CVE. Ubuntu, Debian and Temurin had no fixed release yet: where a fixed build exists, update; where it does not, apply the mitigation.
What has to be true
All four must hold: a HotSpot-based JDK, version 17.0.6 or later and not yet patched, running on x86-64, on a CPU with AVX-512 VBMI. HotSpot-based covers OpenJDK, Oracle Java SE and GraalVM as well as builds such as Corretto, Temurin, Microsoft Build of OpenJDK and Azul Zulu. The OpenJDK risk matrix ticks only 17, 21, 25 and 27, and Red Hat’s bulletin states that OpenJDK 8 and 11 lack the vulnerable code path, although Oracle lists Java SE 8u511b33-perf as affected. Users of 8 and 11 still need the update, because the same advisory includes CVE-2026-114939, CVE-2026-114940 and others that affect every version.
VBMI stands for Vector Bit Manipulation Instructions, part of the AVX-512 family. The researchers and Red Hat list Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, plus AMD EPYC and Ryzen from Zen 4 onwards. Cloud documentation mostly says an instance “supports AVX-512” without naming VBMI, so in the end you have to read the CPU flags on the host itself.
The write-up says the current attack is sensitive to round-trip time and jitter, works differently per product, and that the full chains for individual products will stay private until people have had time to upgrade. The attack surface is Base64 decoding, though, which JWTs, Basic auth headers and plenty of API payloads go through, and most services would struggle to say for certain whether outside input can reach it.
The trigger is one malformed padding character, together with a CPU that happens to support one instruction set, the sort of detail that never makes a meeting agenda. The check is just as small: a CPU flag on each host and the JDK version of each service.
One image, different answers per host
I ran these checks on OrbStack with the official images. They only read the JVM version and flags; nothing sends Base64 input and the vulnerability is not reproduced. OrbStack runs on Apple Silicon here, so linux/amd64 containers go through Rosetta:
# Version and flag defaults, once on amd64 and once on arm64
for plat in linux/amd64 linux/arm64; do
for i in eclipse-temurin:21-jdk amazoncorretto:21; do
docker pull --quiet --platform "$plat" "$i" >/dev/null
docker run --rm --platform "$plat" "$i" java -version 2>&1 | grep 'Runtime Environment'
docker run --rm --platform "$plat" "$i" java -XX:+PrintFlagsFinal -version 2>/dev/null \
| awk '$2 ~ /^(UseBASE64Intrinsics|UseAVX)$/ {print " " $2 "=" $4}'
done
done
Actual output (Docker 29.8.2, 18 November 2026):
OpenJDK Runtime Environment Temurin-21.0.13+9 (build 21.0.13+9-LTS)
UseAVX=0
UseBASE64Intrinsics=false
OpenJDK Runtime Environment Corretto-21.0.13.1.2 (build 21.0.13.1+2-LTS)
UseAVX=0
UseBASE64Intrinsics=false
OpenJDK Runtime Environment Temurin-21.0.13+9 (build 21.0.13+9-LTS)
UseBASE64Intrinsics=true
OpenJDK Runtime Environment Corretto-21.0.13.1.2 (build 21.0.13.1+2-LTS)
UseBASE64Intrinsics=true
The first two are amd64. Rosetta offers no AVX, so the JVM sets UseAVX to 0 and switches the Base64 intrinsic off by itself. The last two are arm64, where the intrinsic is on but uses the ARM64 implementation, which this bug does not touch. 17 and 25 behave the same. What that value is on a real x86 host with AVX-512 I cannot test from here, so the same command has to be run on the host itself.
The JIT looks at the CPU underneath when the JVM starts and only then decides what machine code to generate. That layer is absent from the image and from the SBOM, so the inventory can only be done per host.
Mitigating until you can patch
Both the write-up and Red Hat give two flags, and Red Hat prefers -XX:-UseBASE64Intrinsics, which turns off only the hardware-accelerated Base64 encoding and decoding and leaves other AVX-512 optimisations alone. In practice the two behave very differently:
# -XX:-UseBASE64Intrinsics, Temurin 17, 21, 25, accepted on amd64 and arm64
bool UseBASE64Intrinsics = false {product} {command line}
# -XX:UseAVX=2 on amd64 (Rosetta, no AVX2): warning only
OpenJDK 64-Bit Server VM warning: UseAVX=2 is not supported on this CPU, setting it to UseAVX=0
# -XX:UseAVX=2 on arm64: the JVM refuses to start, exit code 1
Unrecognized VM option 'UseAVX=2'
Error: Could not create the Java Virtual Machine.
UseBASE64Intrinsics is a product flag in the OpenJDK source, so it needs no -XX:+UnlockDiagnosticVMOptions, and the Temurin 17, 21 and 25 builds I tried accepted it directly. UseAVX exists only in x86 JVMs, and the write-up warns against using it anywhere else. A mixed x86 and ARM64 fleet that sets it through one shared JAVA_TOOL_OPTIONS will stop its ARM64 services from starting. Disabling only the Base64 intrinsic directly affects Base64 encoding and decoding; I have not measured the overall cost.
Either way the JVM has to restart, because flags are only read at start-up.
You can pass the flag through JAVA_TOOL_OPTIONS too, but starting a fresh java only proves that new JVM picked it up. To check the service itself after the restart, look at the running process. jcmd ships only with the JDK, and the eclipse-temurin:21-jre image I tried does not have it, so on JRE-only images check the service’s start-up arguments and environment instead:
# Check the flags the running JVMs actually use, rather than starting a new java
for pid in $(pgrep -x java); do
echo "$pid $(jcmd "$pid" VM.flags 2>/dev/null | tr ' ' '\n' | grep BASE64)"
done
In an amd64 container on OrbStack, with a long-running JVM started under JAVA_TOOL_OPTIONS=-XX:-UseBASE64Intrinsics, the actual output was 1 -XX:-UseBASE64Intrinsics. Started without the flag it printed just 1. An empty result is not proof the flag is unset: jcmd can also fail, for example when you run it as a different user from the service.
Once the fixed release is in, take the flag out again. A mitigation left in place becomes a setting nobody can explain a year later, and Base64 encoding and decoding stay on the slower path the whole time.
Where each vendor stands
| Target | Status at time of checking |
|---|---|
| OpenJDK | advisory on 2026-11-17; CVE-2026-114938 affects 17, 21, 25 and 27 |
| Oracle | Critical Security Patch Update on 11-17; fixed releases such as Java SE 17.0.21.1 (17.0.21.1+4) are out |
| Amazon Corretto | 17.0.21.1.2, 21.0.13.1.2 and 25.0.5.1.2 released on 11-17 listing this CVE; the amazoncorretto:17, 21 and 25 images I tested already carry the fix |
| Eclipse Temurin | still on 17.0.21+9, 21.0.13+9 and 25.0.5+8; the support page lists 17.0.21.1 and others as the next release in November, with builds taking up to three weeks and popular platforms usually within days |
| Red Hat | bulletin RHSB-2026-018 rates it Critical and says fixes are being expedited |
| Ubuntu | archive at 17.0.21.0.3, 21.0.13.0.3 and 25.0.5.0.3, all inside the advisory’s affected range |
| Debian | the tracker lists this CVE and marks openjdk-17, 21 and 25 as vulnerable in every suite; bookworm-security 17.0.21+9 and trixie-security 21.0.13+9 and 25.0.5+8 predate the fix |
I did not find primary pages for fixed builds from Microsoft Build of OpenJDK or Azul Zulu, so go by their own notices. Distribution packages often lag by a few days; use java -version to see what is really running and compare it with the vendor’s notice.
Published details, attacks in the wild and KEV
The write-up explains the cause of the out-of-bounds write and the general exploitation technique, and includes a trigger input, so this post does not link to it. Product-specific chains are not public yet, though the researchers say they will publish them later, and Wiz’s blog says it built a PoC within two hours. CISA’s alert of 17 November does not say the flaw has been exploited. The latest KEV catalogue is 2026.11.17, released on 2026-11-17 with 1,782 entries, and it does not include CVE-2026-114938.
Inventory and update
- List everything that runs a JVM: container images, batch jobs, CI runners, and third-party products that ship their own JRE. Put internet-facing services such as Keycloak, Jira and Jenkins first.
- On every x86 host that actually runs a service, check the CPU flags. The hardware condition holds only if
avx512f,avx512bw,avx512vbmiandbmi2are all present; ARM64 hosts can wait. - Run
java -versionwith the Java executable the service actually uses; thejavayour shell finds may not be the one bundled with the service. Only 17.0.6 and later that have not reached this round’s fixed releases (17.0.21.1, 21.0.13.1, 25.0.5.1) are in scope, but 8 and 11 still need the update for the other CVEs in the release. - Where a fixed release exists, switch JDK and restart. If you use official images, make sure the tag now points at the fixed build rather than pulling the same old digest again.
- Where none exists yet, add
-XX:-UseBASE64Intrinsics, restart, and usejcmdto confirm the running service carries the flag. Keep-XX:UseAVX=2off mixed fleets. - Once the fixed release is running, remove the mitigation flag.
# Run on the real host: all four CPU flags must appear; then check the JVM and the intrinsic
grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' | grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'
java -version 2>&1 | grep 'Runtime Environment'
java -XX:+PrintFlagsFinal -version 2>/dev/null | awk '$2 == "UseBASE64Intrinsics" {print $2 "=" $4, $5, $6}'
Run in an amd64 container on OrbStack, the second time with JAVA_TOOL_OPTIONS=-XX:-UseBASE64Intrinsics, the actual output was:
OpenJDK Runtime Environment Temurin-21.0.13+9 (build 21.0.13+9-LTS)
UseBASE64Intrinsics=false {product} {default}
OpenJDK Runtime Environment Temurin-21.0.13+9 (build 21.0.13+9-LTS)
UseBASE64Intrinsics=false {product} {environment}
Rosetta exposes none of the four CPU flags, so the first command prints nothing. {default} in the last column means nobody set the value and the JVM chose it; {environment} or {command line} means someone did. The column says nothing about whether the CPU has VBMI.
Notes
- CVE-2026-114938 affects unpatched HotSpot-based JDK 17.0.6 and later; standard OpenJDK 8 and 11 lack the code but still need the update for the other CVEs.
- It needs x86-64 with AVX-512 VBMI. ARM64 is unaffected, and one image can give different answers on different hosts.
- Oracle and Corretto have fixed releases; Temurin, Ubuntu and Debian did not when I checked.
-XX:-UseBASE64Intrinsicsis a product flag and works on both x86 and ARM64.-XX:UseAVX=2stops ARM64 JVMs from starting, so keep it off mixed fleets.- When I checked there was no report of exploitation in the wild and KEV had not listed it, but CISA has called for immediate patching.
Further reading
- OpenJDK Vulnerability Advisory: 2026/11/17: the OpenJDK advisory, a primary source and the reference for affected versions and CVSS.
- CISA: Critical Java Vulnerability Requires Immediate Patching: CISA’s alert, a primary source with inventory advice.
- Red Hat RHSB-2026-018: Red Hat’s bulletin, a primary source with the clearest statement of conditions and mitigations.
- Oracle Critical Security Patch Update, November 2026: Oracle’s advisory, a primary source for affected Oracle products.
- JDK 17.0.21.1 Release Notes: Oracle’s release notes, a primary source for the fixed version.
- Amazon Corretto 25 changelog: Amazon’s release notes, a primary source; 17 and 21 have their own repositories.
- Adoptium support page: Temurin’s release schedule, a primary source for the next version and build delays.
- Launchpad openjdk-21: Ubuntu’s package publishing history, a primary source; swap the package name for other versions.
- Debian openjdk-21: Debian’s official tracker, a primary source for package versions.
- OpenJDK fix commit: OpenJDK source, a primary source, cited only for the fact that it changes the x86-64 stub.
- CISA Known Exploited Vulnerabilities Catalog: CISA’s official list, a primary source for vulnerabilities confirmed as exploited.
- Wiz: CVE-2026-114938: a security vendor’s blog, a secondary source and where the PoC claim comes from.
- AVX-512 (Wikipedia): an encyclopaedia entry, background on VBMI only.
Ideas and technical judgement by Sheng; drafted with Claude · reviewed by Codex and Grok; JVM versions and flags tested on OrbStack with eclipse-temurin and amazoncorretto 17, 21 and 25 images, vulnerability not reproduced; facts checked up to 08:25 on 18 November 2026 (Taipei time)