Fluxtail
Security

Java4Shell: How OpenJDK’s Base64 Flaw Affects Running Services

Understand Java4Shell (CVE-2026-114938), how it can affect Java servers and clients, and how to verify the fix on each running JVM.

By Fluxtail Engineering

OpenJDK is open-source software that runs Java applications. It includes HotSpot, a Java virtual machine (JVM), and standard libraries that applications use for common tasks. A service might get its Java runtime from a host package, a container base image, or a Java Development Kit (JDK) bundled with the application.

On November 17, 2026, the security research firm Calif disclosed Java4Shell (CVE-2026-114938), a flaw that researcher Quang Luong found and reported in OpenJDK’s Base64 decoder. Base64 represents binary data as text. On affected Java builds, the decoder uses a faster path on x86-64 processors that support a set of instructions called AVX-512 VBMI. Malformed input can make that decoder write past the memory set aside for its output. The resulting damage can crash the Java process, expose data from its memory, or let someone run code with the application’s permissions. Calif says Oracle released its patch at 21:00 UTC that day. Calif’s original disclosure

An application might reach the decoder while doing something routine, such as turning a Base64-encoded file upload back into a file. Calif demonstrated remote code execution in selected applications, including some where the input reached the decoder before login. The firm also reported a client-side demonstration in which a Java program opened an encrypted TLS connection to an untrusted server. That matters for background workers and other Java clients that make outbound connections without accepting incoming requests. Calif’s original disclosure

OpenJDK’s advisory marks HotSpot branches 17, 21, 25, and 27 as affected. Calif traces the upstream code to OpenJDK 17.0.6. Oracle’s own list includes a specific Java 8 build, 8u511b33-perf, and several Oracle GraalVM releases. For any one deployment, exposure depends on the running build, the CPU features the JVM can use, and whether untrusted input reaches the affected decoder. The demonstrations show what was possible in the tested applications. Whether anyone exploited the flaw outside those tests or compromised a particular deployment remains open. OpenJDK advisory · Oracle’s Java risk matrix · Calif’s findings

How to verify the Java4Shell patch on running services

If you run Java services or workers in containers or on VMs, use the running build to check whether each process has its distributor’s fix. Then check that each replacement started and can do its usual work. Whether anyone exploited an older build is a separate investigation.

Identify the running JVM and its CPU features

Use the load balancer or service discovery to list instances that serve requests. Also check your workload and scheduled-job inventory for Java clients and workers that process untrusted data or connect to untrusted servers. Include software that ships its own runtime. For each running instance, record its pod, container, or VM identity and find the process ID of the JVM doing the work.

Use the local process ID with jcmd where the tool is available and local attach is supported and authorized. In this example, PID 42 is the application’s JVM.

jcmd 42 VM.version

Oracle documents VM.version as a low-impact command that reads version information. Run jcmd on the same machine with the same effective user and group IDs as the target JVM. In a container, use an authorized environment that can see and attach to that process. If attach is unavailable, look for trusted runtime inventory or a startup record that identifies the build of that exact process. Mark the build as unverified if you cannot tie it to the running process. Oracle’s jcmd reference

Record the full build and identify the distributor from the JVM output and your package or image records. Running java -version in a shell starts whichever Java executable the shell selects. The application could still be using a different or older JDK. Also record the architecture and whether AVX-512 VBMI is available to the runtime. For a VM, check which CPU features the guest can see. Calif’s affected-build and hardware guidance

Match each build to the vendor advisory and finish the rollout

Check the advisory from the distributor that supplied the JVM. Oracle’s November 17 release notes list security baselines 17.0.21.1+4, 21.0.13.1+4, 25.0.5.1+4, and 27.0.1.1+4 for those Oracle Java release families. Another distributor can ship the fix under a different build or package number. If the application bundles Java, check whether its vendor supplies an application update with a fixed runtime. Oracle’s November 17 release notes

In this illustrative checkout-api rollout, both replicas run on x86-64 with AVX-512 VBMI available. The worksheet pairs a direct check of each running JVM with a separate health result from that replica.

Serving replica Distributor Running build Oracle’s guidance Observed health
checkout-new Oracle JDK 17.0.21.1+4 Matches the November 17 security baseline Started, then returned HTTP 200
checkout-old Oracle JDK 17.0.21.0.3+1 Belongs to the affected 17.0.21.0.3 release Returned HTTP 200 on the affected build

Replace the older replica through your normal rollout process and check the JVM on its replacement. Compare the replicas you checked with the list still serving traffic. For scheduled jobs, check the runtime used by a new run after the update. The image tag or package record tells you what shipped or was installed. The process check tells you what is running. The same distinction applies when checking the running PHP fix. PHP2Shell patch verification

Check startup and health on each replica

Look for a startup record from the replacement, then check a successful request to your usual read-only health endpoint. Use probe results or application logs that identify the responding replica. An HTTP 200 from the load balancer could come from either replica. For a worker without an HTTP endpoint, look for its normal startup and successful work records tied to the replacement instance.

These synthetic rows show what the example rollout might produce. The messages and fields depend on the application and collector.

2026-11-18T00:12:04Z service=checkout-api replica=checkout-new message="application started"
2026-11-18T00:12:15Z service=checkout-api replica=checkout-new message="health request completed" method=GET route=/healthz status=200
2026-11-18T00:12:18Z service=checkout-api replica=checkout-old message="health request completed" method=GET route=/healthz status=200

If these logs already go to Fluxtail, open their stream in Live Tail. If your collection path populates Fluxtail’s Service field with checkout-api, use that filter and search the message text for application started and health request completed in turn. If service=checkout-api arrives only as part of the message, search for checkout-api in the message text and look for the two markers in the matching rows. For retained logs, set the time window to the rollout. Fluxtail Live Tail

Check the replica identity in each matching row. If it’s missing, use probe or deployment records to identify the responder. If an expected row is missing, check whether the application emitted it and whether the collector sent it to the right stream. For clearer startup and request records on future rollouts, see the Spring Boot structured logging guide.

Mark each runtime as patched once its running build matches its vendor’s fix guidance. Record which replacements passed their health checks. Keep evidence about earlier activity with the separate incident investigation.