Java4Shell (CVE-2026-114938): Patch Java's Base64 RCE Flaw Now

Java SE / OpenJDK (Java4Shell) security advisory: Java Base64 RCE, patch now. Critical, CVE-2026-114938. Not in CISA KEV.

Status: CISA urges organisations to patch immediately for "Java4Shell", a memory-corruption flaw in the Base64 decoder of the x86-64 HotSpot Java Virtual Machine (JVM). CISA, Red Hat and Canada's Cyber Centre call it critical. Oracle and OpenJDK score it CVSS 3.1 8.1 (High) because attack complexity is high, and the NVD record shows that same 8.1 High score. CISA says specially crafted Base64 input could crash the JVM and allow unauthenticated remote code execution. The flaw is tracked as CVE-2026-114938. Oracle and the OpenJDK project released fixes on 17 November 2026 in Oracle's November Critical Security Patch Update and the OpenJDK 2026/11/17 advisory. The flaw is not in CISA's Known Exploited Vulnerabilities (KEV) catalog. Canada's Cyber Centre says it is not aware of active exploitation. Red Hat had not yet shipped fixed RHEL or Red Hat build of OpenJDK packages. We checked at 04:39 EST on 18 November 2026.

Executive Summary

Java4Shell is an out-of-bounds write in the hardware-accelerated Base64 decoding routine that HotSpot uses on x86-64 CPUs with the AVX-512 VBMI instruction set. Base64 decoding is everywhere in Java services, and much of it happens before a user logs in: HTTP Basic credentials, SAML artifacts and CSRF tokens are all Base64, as Red Hat's security bulletin notes. That is what makes an unauthenticated attack possible.

Three conditions decide whether a given service is exposed:

  1. A Java runtime with the vulnerable code. The OpenJDK advisory marks Java 17, 21, 25 and 27 as affected and does not mark Java 8 or 11 as affected. Oracle's advisory also lists the Java SE 8 Performance Pack build 8u511b33-perf, plus GraalVM for JDK 17.0.21.0.3 and 21.0.13.0.3 and GraalVM Enterprise Edition 21.3.20.0.3. Red Hat and Calif put the start of the Java 17 range at 17.0.6.
  2. A CPU that reports AVX-512 VBMI. Red Hat lists Intel Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD Zen 4 or newer. ARM64 and older x86-64 chips are not known to be affected.
  3. Attacker-controlled Base64 reaching the JDK's standard decoder.

Scores differ, and the difference matters for triage. Oracle and OpenJDK rate it 8.1 (High) under CVSS 3.1, with high attack complexity and no privileges required. Calif, which found the flaw, scores it 9.2 under CVSS 4.0. Red Hat's preliminary score is 9.8 under CVSS 3.0. It rates the issue Critical. Amazon scores it 8.8 under CVSS 3.1 with low privileges required (PR:L) rather than none. Amazon says that, depending on context, the result is a JVM crash for unauthenticated users or code execution by an authenticated low-privileged user in some environments. CISA and Canada's Cyber Centre call it critical.

What has been shown, and what has not. According to Calif's write-up, as reported by The Hacker News (THN), Calif ran code without a login on default installs of Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate. It also reports code execution with a login on Apache Kafka Broker, Apache Solr, Elasticsearch, Gerrit and Spinnaker, and says these lists are incomplete. Calif is withholding those per-application exploits. THN reports that other labs have reproduced crashes and a memory leak, but not code execution. Datadog Security Labs calls denial of service "the more readily achievable impact" and says reliable code execution needs work tailored to each application and JVM. Tenable also reports that a researcher at Immersive said on 17 November that they had built a working exploit that must be adapted to each application, and named Apache Tomcat Basic authentication and the java-jwt library as ways to reach the flaw before login; Tenable says the post does not say that the exploit has been released. THN reports that Blackpoint Cyber's local JavaBleed test read a secret stored next to the overflow in all 30 runs. How to read 'high attack complexity' (Rescana guidance): AC:H reflects the per-application work needed for code execution, not the crash, which needs no login. Because a triggering input has been public since 9 September (per THN), exposure did not start on 17 November. Plan for crashes first, patch pre-login identity and CI/CD services as if code execution were close, and treat keys and tokens held in an exposed JVM's memory as at risk.

Technical Information

Where the bug is. On AVX-512 VBMI hardware, HotSpot's optimising compiler (C2) replaces java.util.Base64 block decoding with a native "intrinsic". THN notes that Java switches to the intrinsic only after compiling the decoder, which it does for code that runs often, so a quick test on a freshly started JVM can give a false negative (Rescana guidance). Datadog's analysis explains that the routine decoding the final fragment of a malformed, padded input computes a negative output length. That negative length then becomes an all-ones AVX-512 store mask, so a full 64-byte store goes out. In Datadog's test it wrote 62 bytes past the end of the destination array. THN, summarising Calif, puts the overrun at 62 or 63 bytes. The decoder then throws IllegalArgumentException, but only after the write. Catching the exception does not undo the heap damage.

Why "crash" and "RCE" are both right. The overwritten bytes are largely fixed; Datadog describes a repeating pattern. An attacker controls where they land, not what they are. That is why early source readings pointed only to a crash. Calif's reported technique, per THN, uses the overflow to enlarge a neighbouring array. It then uses that array for arbitrary writes and a second array for memory reads, a step Calif calls "JavaBleed". Each target needs its own heap layout and timing, and a failed attempt can crash the JVM. THN reports that an attacker can often try again.

Which decoders. In the OpenJDK 17.0.6 build Datadog tested, the vulnerable path served Base64.getDecoder() and Base64.getUrlDecoder(). Base64.getMimeDecoder() bypassed it. Libraries that use their own Base64 code, or that reject malformed input before calling the JDK, may behave differently. Code review can therefore lower the priority of a service, but only after you trace the actual decode call.

It is not only servers. Calif says, per THN and Tenable, that a Java program can be compromised when it opens a TLS connection to an attacker-controlled server, even if it offers no service. Calif has not published how that path reaches the decoder. Treat this as reported, not confirmed. It still means outbound-only Java processes, such as integration agents, schedulers and connectors, belong in scope.

PoC status (described, not linked). Tenable says Calif published a proof of concept alongside its write-up. It triggers the bug to show whether a given Java build and CPU reproduce it, but does not take control of a system. THN reports that a malformed input triggering the overflow was posted publicly on 9 September in a proposed OpenJDK fix that was closed unmerged in early November. Treat the trigger condition as public knowledge. Application-specific exploit chains had not been released when we checked.

Affected Product Versions

Line Affected (per Oracle / OpenJDK) Fixed build (Oracle)
Java 17 17.0.6 through 17.0.21.0.3 17.0.21.1
Java 21 21.0.13.0.3 and earlier 21.0.13.1
Java 25 25.0.5.0.3 and earlier 25.0.5.1
Java 27 27.0.1.0.3 and earlier 27.0.1.1
Java 8 Performance Pack 8u511b33-perf 8u513 (standard and Performance Pack)
GraalVM for JDK / Enterprise 17.0.21.0.3, 21.0.13.0.3 / 21.3.20.0.3 17.0.21.1, 21.0.13.1 / 21.3.20.1

The fixed-build numbers are from Oracle's release notes as reported by THN and Datadog. Expect no fix for out-of-support releases after Java 17: Amazon lists Corretto 22, 23, 24 and 26 as "No Fix Planned". Treat any such runtime on a VBMI host as exposed and move it to a fixed 21, 25 or 27 build (Rescana guidance). Datadog gives the first fixed builds as +4 (for example, 17.0.21.1+4). The 17.0.6 floor is from Red Hat, Datadog and, per THN and Tenable, Calif. Standard Java 8 and 11 do not contain the vulnerable code.

Downstream distributions, as of our checks:

  • Amazon Corretto: fixed. Amazon Linux lists Corretto 17, 21, 25 and 27 as fixed on 17 November (ALAS entry); THN cites builds 17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2 and 27.0.1.1.4.
  • Red Hat: RHEL 8, 9 and 10 and Red Hat build of OpenJDK 17, 21 and 25 are listed as affected. Red Hat says it is expediting fixes, and THN reports none had shipped. Images based on RHEL or UBI must be rebuilt once updated base images are out.
  • Azul Zulu: fixed in the November update, per THN, in free builds 17.70.201, 21.54.201, 25.38.201 and 27.30.201.
  • SAP SapMachine: THN says new .1 builds were released, but their release pages do not say what they fix.
  • Eclipse Temurin, Microsoft Build of OpenJDK, BellSoft Liberica: THN says none had listed fixed builds when it checked on 18 November.

How to Find Your Exposure

1. Inventory every runtime, not just the ones your CMDB knows about. CISA warns that JDKs and Java runtimes "may not appear as separate entries in software inventories". Sweep for java binaries and release files under application install directories, not only /usr/lib/jvm. Include jlink-trimmed runtimes shipped inside products, JREs embedded in vendor appliances and virtual appliances, and base layers of container images in your registries. Record vendor and full build (for example 17.0.21.0.3 versus 17.0.21.1), as Canada's Cyber Centre advises; the patch-level digit is what separates affected from fixed. Non-Oracle distributions use their own version numbers, so compare each build with its own distributor's fixed version. If you hold SBOMs, query them for JDK and JRE components and for Java base images (Temurin, Corretto, Zulu, UBI OpenJDK and similar tags) by version, but treat a clean result as incomplete: jlink runtimes and appliance JREs are often missing from SBOMs (Rescana guidance).

2. Check the running process, not the shell's java. Datadog recommends jcmd <pid> VM.version and jcmd <pid> VM.flags -all against each live JVM. A host can have a fixed JDK on PATH while the service runs a bundled one. Restart after patching, because an unrestarted process keeps the old runtime in memory. jlink images often lack jcmd. Run it as the user that owns the process, from a full JDK of the same major version, or read the image's release file (Rescana guidance).

3. Check the CPU the JVM actually sees. On Linux, Red Hat's check is grep -o avx512vbmi /proc/cpuinfo | head -1; no output means no AVX-512 VBMI. Tenable, citing Calif, says the vulnerable code needs avx512f, avx512bw, avx512vbmi and bmi2 all present. A broad avx512 match is not enough, because some older Xeons report AVX-512 without VBMI. On any operating system, including Windows, the most direct answer comes from the JVM itself: the CPU line in jcmd <pid> VM.info (and in any hs_err_pid log) lists the features HotSpot detected, so look for avx512_vbmi, and jcmd <pid> VM.flags -all shows UseAVX, where 3 means AVX-512 code paths are on. Run the check inside each VM, because hypervisor CPU models can hide or expose features to guests, and the JVM reads them only at startup, so a VM moved to newer hardware can become exposed at its next restart. Containers see the host's CPU flags (Rescana guidance).

4. Understand why the same image can be safe on one host and exposed on another. HotSpot picks the intrinsic at run time from the CPU it detects. A container image that is harmless on an ARM64 node, or on an older Xeon, becomes exposed when a scheduler or autoscaler places it on an Ice Lake or Zen 4 host. Red Hat names cloud instance families most likely to be affected, including AWS C6i/M6i/R6i and C7i/M7i/R7a, GCP C3/C3D/C4/C4D/N4/N4D/M3/M4, and Azure Dsv5/Dsv6/Dasv6/Dasv7 and related v5-v7 D-series SKUs. Mixed node pools need per-node answers, not a per-image answer.

5. Prioritise by pre-auth Base64 exposure. Patch internet-facing systems first, as CISA advises. Within that group, start with identity and access services and anything that decodes Basic auth, SAML or tokens before login. That includes SSO and IdP servers such as Keycloak and PingFederate, Atlassian Jira and Confluence, CI/CD and artifact systems such as Jenkins and Artifactory, and API gateways written in Java, and any service that uses Apache Tomcat Basic authentication or the java-jwt library before login (per Tenable). Those services are both the most reachable and the most damaging to lose: an IdP or build server compromise reaches far beyond one host.

Workaround and Mitigation

Vendor and government guidance. CISA says that if patching cannot happen immediately, you should restrict and validate Base64 input at the application boundary, enforce input-size limits and limit unauthenticated access. The Cyber Centre adds isolating internet-facing applications from public networks until a fix is applied. Rescana guidance: where you can, put login and admin endpoints behind a VPN or an authenticating reverse proxy, so that unauthenticated traffic does not reach the JVM. SSO in front of an app does not help if the identity provider is itself an affected Java service, and none of these controls stops logged-in users, given Calif's post-login results and Amazon's PR:L score.

Limits of those controls (Rescana guidance). Per Datadog's analysis the malformed input that triggers the write is short, so size caps alone will not stop it. Validating structure at a WAF or reverse proxy is more useful: reject credential and token values where = appears other than in the last one or two positions, or where a padded value's length is not a multiple of four, without breaking unpadded URL-safe tokens such as JWT segments. A WAF only sees the fields it knows to be Base64. It misses values nested inside other encodings (SAML XML, JSON bodies, MIME parts), non-HTTP listeners, traffic where TLS ends at the JVM, and the outbound TLS-client path described above, which needs egress restriction instead. Calif also reports via THN that a proxy that buffers complete requests, as nginx does by default, blocks some attacks but not all.

JVM-flag stopgaps: not in Oracle's or OpenJDK's guidance. Red Hat's bulletin RHSB-2026-018 recommends two startup flags as interim mitigations until a fixed build is installed. They are not part of Oracle's or OpenJDK's advisories; Tenable notes the same. Datadog's tests found that both stopped the out-of-bounds write. THN reports Upwind also found both flags stopped the overflow. Test whichever flag you choose on a non-production instance of each service before rollout, checking that it starts and performs acceptably:

  • -XX:-UseBASE64Intrinsics disables only the accelerated Base64 code, so Base64 decoding falls back to the slower non-intrinsic path while other AVX-512 optimisations stay on. Red Hat calls this the preferred option. Unlike UseAVX, it is a standard option that ARM64 JVMs also accept, so it is the safer choice for mixed-architecture startup profiles (Rescana guidance).
  • -XX:UseAVX=2 limits the JVM to AVX2 and so disables all AVX-512 code paths, not only Base64. Datadog advises assessing the performance impact on your workload. Canada's Cyber Centre also suggests this flag. Apply it only on x86-64: Red Hat and Datadog warn that ARM64 JVMs do not recognise it and will fail to start, so never push it through a shared, architecture-agnostic startup profile.

If you deploy either flag through JAVA_TOOL_OPTIONS or a launcher config, confirm it took effect with jcmd <pid> VM.flags -all. Services started by systemd units, init scripts or appliance supervisors may not inherit the environment you set (Rescana guidance). THN says Upwind warns that -XX:DisableIntrinsic=_base64_decodeBlock on its own stops Java from starting. That is expected, because DisableIntrinsic is a diagnostic option that HotSpot accepts only after -XX:+UnlockDiagnosticVMOptions. Do not use it as a workaround; -XX:-UseBASE64Intrinsics is a standard option that closes the same path (Rescana guidance). For products that bundle their own JVM, Red Hat advises applying the flag or the update through the product vendor's supported path rather than swapping the JVM yourself. Remove the flags once fixed builds are in place.

Kubernetes placement (Rescana guidance). Where node-feature-discovery is installed, VBMI-capable nodes carry the label feature.node.kubernetes.io/cpu-cpuid.AVX512VBMI=true. Until a workload is patched, a required node-affinity rule with the DoesNotExist operator on that label keeps it off those nodes, and moving it to ARM64 nodes also works if the image is built for arm64. Affinity applies at scheduling, so running pods must be restarted to move, and it is a stopgap until fixed builds are in place. Treat the label as a scheduling aid, not proof, and confirm with the JVM's own view (step 3 under How to Find Your Exposure).

Indicators of Compromise

Oracle, OpenJDK, CISA and Canada's Cyber Centre have published no indicators of compromise, and no exploitation has been reported to derive them from. The items below are detection leads, not IoCs. CISA says to monitor for rogue runtimes, unexpected JVM crashes and suspicious Java processes. In practice:

  • Crash artefacts. New hs_err_pid*.log files, core dumps, or bursts of service restarts from systemd, Kubernetes (CrashLoopBackOff, rising restart counts) or appliance watchdogs on VBMI-capable hosts. Datadog observed the corruption surfacing later as a crash during garbage collection, so the crash stack may not point at Base64 at all. hs_err_pid files are written to the process working directory (or the temporary directory if that is not writable) unless -XX:ErrorFile points elsewhere. In containers they are lost when the container is replaced, which Kubernetes does on every restart, so point -XX:ErrorFile at persistent storage (for example a mounted volume path ending hs_err_pid%p.log) or collect it with your logs. The CPU section of each file shows whether the host reported avx512_vbmi. Correlate crash times with access logs for malformed Base64 in the minutes before, and look back to 9 September, when a trigger was posted publicly, not only to 17 November (Rescana guidance).
  • Decoder errors. THN reports Upwind's suggested signals: bursts of IllegalArgumentException from java.util.Base64$Decoder, and Base64 values of 32 characters or more in Authorization headers that end in = but whose length is not a multiple of four. Upwind warns that none of these proves compromise, and their absence does not prove safety.
  • Volume patterns. THN reports that Blackpoint Cyber crashed a test Spring Boot app by sending about 28,300 malformed login requests in roughly 20 seconds without logging in. Alert on failed-auth spikes against Java login endpoints.
  • Post-exploitation. Watch for shells, script interpreters, curl or wget spawned by a JVM process, new outbound connections from Java services, new JARs or web-shell-like files in application directories, and unexpected Java processes or runtimes (Rescana guidance). On Windows, the equivalent is cmd.exe or powershell.exe children of java.exe. Baseline first: build servers and some admin tools legitimately spawn shells from a JVM, so treat a new child process as a lead, not proof. A JavaBleed-style memory read leaves no host artefact, so crash-log and access-log correlation remains the more reliable early signal. Canada's Cyber Centre advises reviewing web server, application, authentication and endpoint logs.

Third-Party Risk (TPRM) Angle

For many organisations, much of the Java4Shell exposure sits outside their own estate. Java runs inside security and network appliances, storage and backup consoles, monitoring tools, and the back ends of SaaS platforms. Suppliers also run the exact applications Calif named, such as Jira, Confluence, Jenkins, Artifactory, Keycloak and PingFederate, on your behalf. CISA and Canada's Cyber Centre both call out runtimes bundled with third-party products and appliances. Red Hat says products that bundle their own JVM must be fixed through the vendor. You cannot patch those yourself; you can only find them and press the owner.

The risk concentrates in three supplier types:

  • Identity and access providers (hosted Keycloak, PingFederate, SSO brokers). A pre-auth compromise there can mean token or credential theft, and, through JavaBleed-style memory reads, signing material, affecting every app that trusts them.
  • Software supply-chain suppliers (hosted Jenkins, Artifactory and similar). Code execution on a supplier's build or artifact system is a path to tampered software reaching you.
  • Appliance and on-prem product vendors. Fixes arrive on the vendor's schedule, not Oracle's. Distributions without fixed builds when we checked, such as Red Hat's, can hold up downstream products built on them.

Send these three questions to suppliers in those three groups now:

  1. Which of your services that face the internet or handle our data run Java 17.0.6 or later, 21, 25 or 27 below the November 2026 fixed builds (or your distribution's equivalent), an out-of-support Java release after 17, GraalVM, or the 8u511b33-perf build, on x86-64 hosts whose JVM reports AVX-512 VBMI, including inside appliances and containers?
  2. Are any internet-facing Java services that you run for us, or that hold our data, reachable without login?
  3. When will you be on fixed builds or your distribution's fix, and what compensating controls are you using until then?

Ask for evidence, not reassurance: the runtime vendor and full build number of affected services (an SBOM or runtime-inventory extract will do), confirmation that processes were restarted and images rebuilt, and, if they rely on JVM flags, confirmation from jcmd output that the flag is active, the result of a review of JVM crash logs back to 9 September, whether they will rotate identity-provider signing keys and session secrets if they find crash evidence, and a commitment to tell you within an agreed window if they find signs of exploitation. Prioritise suppliers with SSO, CI/CD or data-processing roles. Record answers against each vendor record and set a follow-up for Red Hat-based suppliers, whose fixes were still pending.

What to do internally: inventory and fingerprint runtimes and CPUs, patch pre-auth internet-facing services first, apply a tested interim flag from Red Hat's guidance where you cannot patch today, and alert on JVM crash loops. We will update this advisory if CISA adds the flaw to KEV, exploitation or a public RCE exploit is confirmed, or Red Hat and other distributions ship fixes.

Need help finding which of your suppliers run exposed Java services and tracking their answers? Talk to an expert

Sources

Talk to an expert

Talk to Rescana about this advisory — or about how we help TPRM teams keep reviews moving.