On 17 November 2026, Calif disclosed Java4Shell (CVE-2026-114938), a critical memory-corruption flaw in OpenJDK’s Base64 decoder that allows remote code execution on x86-64 CPUs with AVX-512 VBMI. Oracle fixed it the same day in its November Critical Security Patch Update. Quang Luong of Calif.io found it with OpenAI’s Codex in about an hour, according to Calif.

This OpenJDK Base64 vulnerability is awkward for a reason that has little to do with severity. The flaw lives in the JDK, not in a library in your lockfile, so a lockfile-based dependency scan will not list the JDK that runs your application. Whether a given JVM is exposed also depends on the processor underneath it, so the inventory has to start from the runtime and the CPU, not from the application.

What the flaw needs before it fires

The x86-64 Base64 intrinsic miscalculates the store mask for the padded tail of the input, which causes an out-of-bounds write. Qualys puts it at 62 bytes past the end of the buffer, before Java rejects the input. Endor Labs notes the vulnerable code only runs once C2 has compiled the decoder as hot code.

The reported trigger is 64 “A” characters followed by “=”; older OpenJDK 17 needs 128. Datadog found that the AVX-512 path covers getDecoder() and getUrlDecoder(), while getMimeDecoder() bypasses it. Use that for triage, not as a reason to delay.

It is not a universal exploit. An attacker needs an application-specific way to get Base64 data to the decoder. Calif claims unauthenticated RCE in default configurations for Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate, and authenticated RCE for a Kafka broker, Solr, Elasticsearch, Gerrit and Spinnaker.

Calif published a PoC of the memory corruption and withheld the weaponised exploits. As of 17 November, Upwind reported the flaw was not in CISA KEV and that no exploitation in the wild had been seen. The flaw did leak before the patch, so treat that status as a snapshot.

How to find vulnerable JDK versions

Start with the runtime, including every JDK bundled inside a product or container image, because those never appear in your build files. The affected range, per the sources we read:

  • OpenJDK 17.0.6 and later, plus distributions built from that code (Oracle, Corretto, Temurin, Microsoft, Zulu, Red Hat). Per Endor, 17.0.0 to 17.0.5 are not affected.
  • The lines with fixed releases listed below (17, 21, 25 and 27).
  • OpenJDK 8 and 11 are not affected, according to Red Hat, with one exception: Oracle’s Java 8 performance build (8u511b33-perf).

Datadog lists the fixed releases as 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. Qualys reports Amazon had released Corretto 17.0.21.1.2 with the fix as of 17 November. Red Hat’s bulletin, RHSB-2026-018, was still marked “Ongoing” and lists RHEL 8, 9 and 10 and its OpenJDK 17, 21 and 25 as affected.

Fix status for Temurin, Microsoft, Zulu, Liberica, SapMachine and distro packages was unconfirmed on 17 November. Check your vendor’s advisory before you mark a host done. Fixed versions for Java 8 and 11 are reported inconsistently, so do not plan around them.

Which hosts expose AVX-512 VBMI to the JVM

Calif says the bug needs an x86-64 CPU with AVX-512 VBMI. Red Hat names Intel Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD Zen 4, Zen 5 and Zen 6. ARM, including Graviton, and Skylake or Cascade Lake Xeons are reported to be off the path, though that comes from Equixly, not a primary source.

Run the check inside the guest or container, not on a spec sheet. A virtual machine can mask AVX-512 even when the hardware has it, and most Linux distributions show vbmi in the guest only if the hypervisor passes the flag through. Several vendors check four flags: avx512f, avx512bw, avx512vbmi and bmi2.

grep -owE 'avx512f|avx512bw|avx512_?vbmi|bmi2' /proc/cpuinfo | sort -u
java -XX:+PrintFlagsFinal -version | grep -E ' (UseAVX|UseBASE64Intrinsics) '
jcmd <pid> VM.flags

The first line is the CPU check; kernels spell the flag both avx512vbmi and avx512_vbmi. The second reads the JVM’s own view. Endor’s rule: the path is live if vbmi is present, UseAVX is above 2 and UseBASE64Intrinsics is true. For a process that is already running, jcmd <pid> VM.flags shows what it started with.

The question is not which applications use Base64. It is which processes can see vbmi.

We run Cloud on KVM, so an instance sees only the flags the hypervisor passes through. That holds for a VPS, VDS, MEM or MNG alike: run the grep in each instance and record the answer per instance. On Bare Metal, run the grep on every server, Entry through High end, and record whether vbmi appears.

On MNG, do not assume the management covers the JDK under your application. Confirm with us which software on your instance is in scope. Kafka brokers appear on Calif’s authenticated list. Patching is on record for the managed brokers we run under Event Streaming, but the status of the JDK under them for this flaw is something to confirm with us. For a Kafka you run yourself on a VDS or a Bare Metal server, the JDK is yours to patch.

Sequencing the patch and the UseAVX=2 workaround

Calif lists two stopgaps: -XX:-UseBASE64Intrinsics and -XX:UseAVX=2, which disables AVX-512. BleepingComputer is clear that both are temporary and that updating is still required. Qualys and Red Hat prefer the narrower -XX:-UseBASE64Intrinsics, because UseAVX=2 switches off a wider set of optimisations. We found no published figure for that cost, so measure it on one canary before you roll it out.

The UseAVX=2 workaround has a second catch for mixed fleets. It belongs on x86-64 only: OpenJDK on ARM does not recognise the option and fails to start. Scope the flag by architecture, not by application.

A sequence that holds up:

  1. Inventory every JVM: CPU flags in the guest, JDK version and vendor, and the two flag values. Mark each as live only if all three Endor conditions are true.
  2. Patch the live JVMs first where a fixed build exists for your distribution, starting with internet-facing applications on Calif’s unauthenticated list, then the authenticated ones, then the rest.
  3. Where no fixed build exists yet, apply -XX:-UseBASE64Intrinsics, and fall back to UseAVX=2 only on x86-64 hosts where you need it.
  4. Confirm with jcmd <pid> VM.flags after the restart, then remove the stopgap when the patch lands.

Both stopgaps and the patch need a JVM restart, so you get one restart window per service, not two. If a fixed build is available, restart straight into it and skip the flag. Hosts that show no vbmi stay on the patch schedule, not the emergency one.

When hardware fails under a Cloud instance, the workload moves to another node, usually without the customer noticing. Re-run the grep after any host move, and keep the result next to the instance in your inventory.

Keep that inventory after this ticket closes. The next JDK flaw that depends on a CPU feature will ask the same questions, and you will already have the answers per host.

Sources