A heap overflow in Java's AVX-512 Base64 decoder, reachable before login. Calif reports remote code execution on Keycloak, Jira, Confluence, Jenkins and more. Upwind reproduced the memory corruption in our lab.
java -version on the host or inside the container.Oracle's November 17 Critical Patch Update fixes the bug. The same version number without the .1 suffix is still vulnerable. Downstream builds (Temurin, Corretto, Zulu, and others) are rolling out; check your vendor.
17.0.6 – 17.0.2117.0.21.1Affectedall before 21.0.13.121.0.13.1Affectedall before 25.0.5.125.0.5.1Affectedall before 27.0.1.127.0.1.1Affectedall—Not affectedOpenJDK 8See Oracle advisoryCheck Oracle-perf build as affected. If you run Oracle Java SE 8, confirm against Oracle's page.Sources: Oracle CPU Nov 2026 · OpenJDK advisory · Calif write-up
We keep our own lab results separate from what researchers report, so you know how much weight each claim carries.
-XX:-UseBASE64Intrinsics blocks the bug on every affected versionEvery Java 17/21/25/27 runtime, including JREs bundled in containers and vendor products like Keycloak, Jira or Jenkins.
Internet-facing login and API services on AVX-512 x86 hosts first. They decode Base64 from strangers by design.
Move to the .1 builds. If you can't yet, apply the stopgap:
-XX:-UseBASE64IntrinsicsHunt for malformed Base64, decoder exceptions and JVM crash loops. See detection below.
JAVA_TOOL_OPTIONS, it causes an outage:
-XX:DisableIntrinsic=_base64_decodeBlock Error: VM option 'DisableIntrinsic' is diagnostic and must be enabled via -XX:+UnlockDiagnosticVMOptions.Use
-XX:-UseBASE64Intrinsics instead. -XX:UseAVX=2 also works, but only on x86-64: on ARM the JVM won't start with it.
Three signals from our lab, strongest first. Legitimate padded Base64 always has a length that is a multiple of 4, so the first rule is low-noise.
A token in headers or body that contains = but whose length isn't a multiple of 4. Strip the auth scheme first (Basic, Bearer…). Flag anything that doesn't match:
^(?:[A-Za-z0-9+/_-]{4})*(?:[A-Za-z0-9+/_-]{2}==|[A-Za-z0-9+/_-]{3}=)?$Fired on every trigger attempt in our tests. It's thrown after the overflow, so catching it doesn't prevent the corruption.
IllegalArgumentException: Input byte array has wrong 4-byte ending unit
SIGSEGV in the garbage collector, new hs_err_pid*.log files, and services restarting in a loop. Don't rely on this alone: crashes depend on heap layout.
SIGSEGV … G1ParScanThreadState::copy_to_survivor_space
Details leak ahead of Oracle's patch. CVE-2026-114938 reserved.
Upwind reproduces the 62-byte heap overflow on Ice Lake hardware; confirms 17.0.6+, 21, 25 affected; 11 and ARM clean.
Calif publishes its write-up and RCE claims across 11 products. Exploit chains held back.
Patch available: 17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1. Downstream builds rolling out.
We'll update this page as vendors ship builds and as RCE claims are confirmed or refuted.
Upwind shows which runtimes run an affected JDK, on which CPU, and whether they're reachable from the internet, with runtime detection for post-exploitation behavior.