Java4Shell: Base64 flaw in the JVM puts servers at risk

On November 17, 2026, Oracle and the OpenJDK project released fixes for CVE-2026-114938, a memory-corruption flaw in the HotSpot Java Virtual Machine (JVM) that the research firm Calif, which reported it, calls Java4Shell. A malformed Base64 string, as short as 64 "A" characters followed by a single =, can make the JVM write past the end of a buffer. CISA warned the same evening that specially crafted input "could crash the JVM and allow for unauthenticated remote code execution" and urged organizations to patch immediately. The fixed releases are Java 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1.
The trigger is narrow, the reach is wide. An attack needs an affected Java runtime (OpenJDK 17.0.6 or later, including the 21, 25 and 27 lines), an x86-64 processor with the AVX-512 VBMI instruction set, and an application that decodes Base64 from untrusted input, which, according to Calif, ordinary features such as authentication and file uploads can do. Red Hat lists the AWS, Google Cloud and Azure instance families most likely to qualify. As with PHP2Shell in October, the hard part for most companies is finding every runtime they operate, including the ones hidden inside products.
What happened
Calif, which also took part in finding PHP2Shell, says it used AI to find and exploit the flaw: "It took our own Quang Luong an hour to discover the vulnerability, and two more hours to develop the first working exploit. The total token cost was less than $200." Oracle and Ubuntu credit both Quang Luong and Albert Yang. Albert Mingkun Yang, an OpenJDK contributor, had described the same bug in a public OpenJDK pull request on September 9 without labeling it a security issue, and he is a co-author of the fix.
The disclosure did not go as planned. Oracle had pre-announced a Java security update for Tuesday, November 17, without naming the flaw. That day, Korean media reported that South Korea's National Intelligence Service had warned public institutions to prepare for the announcement of a Java vulnerability, and the CVE number was circulating online hours before any fix. At 16:03 UTC Calif wrote on X: "Since the vulnerability has leaked: Calif discovered and reported Java4Shell." Oracle's update, the OpenJDK advisory and the fixed builds followed at 21:00 UTC, and Calif published its technical write-up minutes later.
What the flaw does
Base64 turns binary data into text, and Java applications decode it constantly: in HTTP Basic authentication headers, JSON Web Tokens, SAML messages and cookies, according to CERT-EU. On x86-64 processors with AVX-512 VBMI, the JVM by default hands most of the decoding to a hand-written machine-code routine once the decoder is used often enough. According to Calif, that routine mishandles a stray = at the end of the input: for the 65-character trigger it computes a negative output length and writes a full 64-byte block, 62 bytes past the end of the 50-byte array Java allocated. Java then rejects the string as invalid, but "by then neighboring objects may already be corrupted", Calif writes.
The attacker controls little of what is written, mainly where it lands. Calif says it turned that into a leak of heap memory and, after shaping the memory with a series of requests, into remote code execution. It also describes a client-side variant: "A Java program that opens a TLS connection to a malicious server can be compromised, even if it exposes no service of its own." Oracle rates the flaw 8.1 (High) under CVSS 3.1 and marks its attack complexity as high; Calif rates it 9.2 (Critical) under CVSS 4.0.
Who is affected
Three conditions have to come together. First, the runtime: OpenJDK 17.0.6 and later and the Java 21, 25 and 27 lines, including builds from Oracle, Amazon Corretto, Eclipse Temurin, Microsoft, Azul and Red Hat, and Oracle GraalVM for JDK. Standard Java 8 and 11 do not contain the vulnerable code, according to Red Hat, but Oracle lists its Java 8 Enterprise Performance Pack (8u511b33-perf) as affected. Second, the processor: Calif names Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD EPYC and Ryzen chips from Zen 4 on. Processors without AVX-512 VBMI use a different decoding path, and ARM64 servers are not affected. Third, the application has to pass attacker-controlled Base64 to the JDK's own decoder.
Calif says it achieved remote code execution without authentication in default configurations of Keycloak, Atlassian Jira and Confluence, JFrog Artifactory, Jenkins and PingFederate, and with authentication in Apache Kafka, Apache Solr, Elasticsearch, Gerrit and Spinnaker, adding: "The lists are non-exhaustive." It is withholding those exploits "until everyone has had a chance to upgrade". By November 18, Atlassian, the Jenkins project, Keycloak and Elastic had published guidance for their products.
Java also ships inside products that few teams count as Java servers. Cisco listed Catalyst Center, Secure Network Analytics and other products as affected on November 18, with no workarounds and fixes still pending, and Siemens advised customers to shut down Industrial Edge Management instances or disconnect them from the internet. CISA warns that "JDKs and Java runtimes may not appear as separate entries in software inventories."
How attacks like this unfold
Crashing a server is far easier than taking it over. Datadog Security Labs writes that "Denial of service is the more readily achievable impact where attacker-controlled input reaches the vulnerable decoder", and VulnCheck found that a single request usually does not crash a server, while a burst of requests reliably brought the service down. Code execution needs an attack chain built for each application; in Calif's words, "there is no universal exploit". Within hours of the release, VulnCheck made a full exploit available to its paying customers, and a researcher at Immersive reported a working, application-specific exploit. Neither was made public.
As of November 18, no successful attack had been confirmed. CISA had not added the flaw to its Known Exploited Vulnerabilities catalog, and Germany's Federal Office for Information Security (BSI), Japan's JPCERT/CC and Canada's Cyber Centre reported no known attacks. Scanning had started, though. GreyNoise recorded exploitation attempts, and Cloudflare's threat intelligence team reported on November 18 that one address first probed a target with an obvious java4shell-* user agent and, about 17 hours later, sent the exploit trigger to the same target "posing as Chrome".
Calif says of its product exploit chains: "We will publish them later." Until then, attackers have to build their own, and the probing shows that some have started.
Why leaders should care
Governments treated the flaw as urgent. On November 17, CISA, with Canada's Cyber Centre, told organizations to look for Java runtimes "bundled with third-party applications or included in container images and appliances". On November 18, CERT-EU, the BSI and Luxembourg's CIRCL published advisories, and Slovenia's government information security office issued a binding order for the organizations under its information security law: identify affected systems within 24 hours, and patch internet-facing ones within 48 hours and the rest within seven days of receiving the order, or of a fix becoming available where none existed yet. In South Korea, the science ministry and KISA urged organizations on November 18 to check and update their systems.
For a company, the exposure is rarely one application. Java runs identity providers, ticketing and wiki systems, build servers, search clusters and message brokers, and many vendor products and appliances carry their own runtime. Each of them waits for its vendor's update: Eclipse Temurin was still rolling out fixed builds on November 18, and one of its maintainers says the Eclipse Foundation is not in the OpenJDK Vulnerability Group, so Temurin starts building only when the fixes become public. A host that is safe today can also become exposed when a resize, a migration or autoscaling moves the workload to a different processor, as Censys and CIRCL point out.
A compromised identity provider or build server can expose credentials, customer data and the software you ship, and the GDPR, NIS2 and DORA obligations that follow an incident apply whether the flaw sat in your code, in the Java runtime or in a supplier's appliance.
What to do now
Find every Java runtime, including the bundled ones
List servers, container images, build agents, appliances and vendor products, and check the runtime each one actually uses:
java -versionwith that runtime, orjcmd <PID> VM.versionfor a running process. For bundled runtimes, ask the vendor.Check which hosts run the vulnerable code
On Linux,
grep -o -w -m1 avx512vbmi /proc/cpuinfoshows whether a host supports the instructions; Calif's own check also requiresavx512f,avx512bwandbmi2. Use the result to set priorities, not to close findings.Update Java and restart every JVM
Install
17.0.21.1,21.0.13.1,25.0.5.1or27.0.1.1, or8u513-perffor Oracle's Java 8 Enterprise Performance Pack, or your vendor's fixed build, and rebuild container images on patched bases. Internet-facing systems come first. Running JVMs keep the old code until they restart. Java 22, 23, 24 and 26 get no fix, so move those workloads to a supported release.Use the JVM flag only as a bridge
Where you cannot update today, Calif and Red Hat point to
-XX:-UseBASE64Intrinsics, which switches the vulnerable routine off, for example throughJAVA_TOOL_OPTIONS, followed by a restart. Confirm it withjcmd <PID> VM.flags -all. The alternative,-XX:UseAVX=2, works only on x86-64: a JVM on ARM fails to start with it. Neither flag appears in Oracle's or OpenJDK's advisory, and neither replaces the update. Cloudflare's web application firewall rules run in log mode by default and cannot see outbound TLS or non-HTTP traffic.Watch for crashes and suspicious Base64
According to the BSI, no concrete indicators of attack were available on November 18, but the signals are known. Look for unexpected JVM crashes and new
hs_err_pidlog files, bursts ofIllegalArgumentExceptionerrors from the Base64 decoder, and requests, for example in the Authorization header, carrying a long Base64 value that ends in a short tail with=, like the published triggers of 64 or 128 "A" characters followed by=.Press your suppliers and test what attackers can reach
Ask every vendor whose product runs on your network or holds your data whether it ships affected Java on affected hardware, and when it patched. Then check your exposure from the outside. Autonomous Penetration Testing runs a full penetration test of the scope you authorize, such as your internet-facing URLs, and reports verified findings within hours.
Pressing questions
We do not develop in Java. Can we still be affected?
Yes. Jira, Confluence, Jenkins, Keycloak, Elasticsearch and Kafka all run on Java, and Cisco and Siemens list products that bundle it. Ask your vendors whether their products ship an affected runtime.
Is Java4Shell being exploited?
As of November 18, no successful attack had been confirmed and the flaw was not in CISA's exploited catalog. Scanning and exploit-shaped requests had begun, and working exploits exist in private hands. Treat every unpatched internet-facing system that meets the three conditions as at risk.
Our servers run on ARM or older processors. Are we safe?
From this flaw, mostly: ARM64 servers and processors without AVX-512 VBMI do not run the vulnerable routine. Update anyway. Workloads move between hosts, and Oracle's November update fixed 11 Java SE vulnerabilities, not only this one.
Is the JVM flag enough?
No. It removes the vulnerable path only from JVMs that were restarted with it, it is not part of Oracle's or OpenJDK's guidance, and the AVX variant stops Java from starting on ARM. Use it to buy days, then update.
Is it safe to test our servers with an exploit?
Not against production. The easiest outcome of a Java4Shell request is a crashed JVM, so exploit code can take down the service it is meant to test. To check a JDK and processor combination, Calif published a small trigger program that writes only inside a real array; it uses the 65-character trigger, so it can miss Java 17 builds that need 128 "A" characters. For an outside view of what attackers can reach, use authorized testing built for production: Cyberware's Autonomous Penetration Testing uses non-destructive techniques, avoids denial-of-service payloads and reports only findings with verified proof of exploit.
Key takeaways
Java4Shell is a flaw in the runtime beneath many enterprise applications, not in any one of them, and it matters wherever an affected JVM, an AVX-512 VBMI processor and untrusted Base64 meet. On November 18 there were no confirmed attacks, crashing a server was far easier than taking it over, and the known working code-execution exploits for real products were private. Calif's hour of AI-assisted research, a month after PHP2Shell, shows how fast the next one can come.
Update every Java runtime and restart it, use the JVM flag only where you cannot update today, and ask your suppliers the same questions. Then verify from an attacker's point of view: Autonomous Penetration Testing shows what is reachable and exploitable on your internet-facing systems.