Java has a new bug with a nickname. Calif Research has disclosed what it calls Java4Shell, tracked as CVE-2026-114938. It is a memory-corruption flaw in the code Java uses to decode Base64 text on processors that support Intel's AVX-512 vector instructions. Oracle published the CVE on November 17 and shipped fixed Java SE builds the same day, as part of a November update that covers 652 vulnerabilities across its products, according to CyberScoop.

The name recalls Log4Shell, the 2021 logging-library disaster. Most of the comparison is marketing. The part that holds up is that Base64 decoding, like logging, runs in a great deal of Java server software, often on data that arrives straight from the network before anyone has logged in. The details are still disputed in places, and those details decide how fast you need to move. This explainer walks through them.

What is the bug?

Base64 is a way of writing binary data as plain letters, digits and a few symbols. Web tokens, authentication headers, email attachments, certificates and many API payloads use it. Every Java application has a Base64 decoder in the standard library.

For speed, OpenJDK uses hand-written processor instructions to decode Base64 when the CPU supports AVX-512, an instruction set found on many x86 server chips. The Hacker News reports that this code runs only on x86-64 processors with AVX-512 VBMI support, when untrusted data reaches the basic or URL decoder of java.util.Base64. The National Vulnerability Database (NVD) record files the issue under Oracle's "Compiler" component.

CyberScoop reports that a short, specially formed Base64 string can make the optimized routine write 64 bytes into a buffer that may hold as few as 50. Writing past the end of a buffer is a classic memory-safety failure. On its own the corruption usually crashes the program. Calif says an attacker who carefully arranges what sits in the surrounding memory can turn it into control of the server. ZeroDayFeed is not reproducing the triggering input.

How bad is it? It depends who you ask

The severity scores already disagree:

  • Calif Research rates the flaw 9.2 out of 10, which is critical. The Hacker News describes this as a CVSS 4.0 score.
  • Oracle's NVD record rates it 8.1 under CVSS 3.1, which is high. The vector is network-reachable (AV:N) but marked high attack complexity (AC:H). Oracle calls it "difficult to exploit" but reachable by an unauthenticated attacker over the network, for example by supplying data to APIs through a web service.

The bigger dispute is about what an attacker can actually do. CyberScoop reports that Calif reported unauthenticated remote code execution in widely used products including Keycloak, Jira, Confluence, Jenkins and PingFederate, and demonstrated a working break-in against Keycloak. Calif also reported authenticated code execution in Kafka, Solr and Elasticsearch, meaning an attacker would first need a valid account. Those are Calif's own research results. Calif says there is no universal exploit and each target needs tailored steps. None of our sources show the vendors of those products independently confirming the results.

Other researchers were more cautious early on. The Hacker News reported that analyses from Upwind and Equixly pointed to a crash of the Java virtual machine, a denial of service, as the likely result, because the bytes written past the buffer are fixed rather than chosen by the attacker. Neither firm reported reproducing the overflow on AVX-512 hardware. Calif's code-execution claims came with its disclosure, and that early disagreement is why Oracle's "high complexity" rating and Calif's "critical" rating can both exist.

For defenders, the practical point is this: even if the result were "only" a crash, a remote attacker who can kill your identity provider or CI server with one request has a problem worth patching. If Calif's code-execution results hold up broadly, the urgency goes up. Gal Elbaz of Oligo Security told CyberScoop that not every Java application is vulnerable and exploitation will not be straightforward, because attackers must account for specific hardware, runtime configurations and application behavior, but that the technical barriers look surmountable.

Is anyone exploiting it?

Not that anyone has reported. As of CyberScoop's publication, no exploitation in the wild had been reported, and none of the related CVEs had been added to CISA's Known Exploited Vulnerabilities (KEV) catalog. CISA published a notice calling for a patch. The NVD entry was still in "Received" status when we checked, which means NVD analysts have not yet added their own assessment to Oracle's submission.

That could change. CyberScoop quotes researchers who say the likelihood of real-world exploitation is high, especially given how fast attackers can use AI.

Found by AI in about an hour

How the bug was found is part of why it is getting attention. According to CyberScoop, Calif said researcher Quang Luong found it in about an hour while auditing OpenJDK with OpenAI's Codex, spending under $200 in computing tokens, and that AI agents produced most of the group's exploit chains nearly on their own. CyberScoop's headline calls it "a vulnpocalypse bug worthy of attention."

The lesson for defenders is less about this CVE than about the trend. Hand-optimized code paths in core runtimes, which few people have reviewed line by line, can now be audited cheaply, and patch cadence needs to keep up.

Which Java versions are affected?

This is where the sources take the most care to read correctly.

The Hacker News reports that the vulnerable code exists in Java 17.0.6 and later, and in Java 21, 25 and 27. It is not in standard Java 8 or Java 11 builds. CyberScoop reports that only systems with AVX-512 are exposed and that ARM64 systems are unaffected; The Hacker News likewise says ARM servers do not run the flawed code.

Oracle's NVD record lists these affected products:

  • Oracle Java SE 8u511b33-perf, 17.0.21.0.3, 21.0.13.0.3, 25.0.5.0.3 and 27.0.1.0.3
  • Oracle GraalVM for JDK 17.0.21.0.3 and 21.0.13.0.3
  • Oracle GraalVM Enterprise Edition 21.3.20.0.3

The Java 8 entry is a specialized "-perf" build, not the standard Java 8 line, which is consistent with The Hacker News's statement that ordinary Java 8 does not contain the code.

Separately, Oracle's advance notice covered 11 Java SE vulnerabilities in the November update. The Hacker News reports that Oracle described all 11 as remotely exploitable without authentication, with a maximum CVSS 3.1 score of 8.1, and did not disclose their CVE identifiers. The Hacker News said there was no evidence at the time that Java4Shell was among those 11, so treat them as a reason to update even Java lines that Java4Shell does not touch.

What It Means for Defenders

Who is affected: Anyone running Java 17.0.6+, 21, 25 or 27 (Oracle Java SE, GraalVM, or OpenJDK-based runtimes that carry the same code) on x86 hardware with AVX-512, especially internet-facing services that decode Base64 input from untrusted users. The Hacker News notes that Oracle's notice does not establish the patch status of third-party OpenJDK distributions, so check your vendor. Treat SSO and identity servers (Keycloak, PingFederate), collaboration tools (Jira, Confluence) and CI/CD servers (Jenkins) as top priority, since those are where Calif reports unauthenticated code execution.

What to patch: Move to the November replacement builds. The Hacker News lists the planned builds from Oracle's ops.java site as:

  • Java 8: 8u513
  • Java 11: 11.0.33.1
  • Java 17: 17.0.21.1
  • Java 21: 21.0.13.1
  • Java 25: 25.0.5.1
  • Java 27: 27.0.1.1

Update Java 8 and 11 too. Java4Shell is not reported to affect standard builds of those lines, but Oracle's November update fixes 11 remotely exploitable Java SE flaws.

Stopgaps if you cannot patch yet: Upwind and Equixly named two JVM options that keep the flawed code from running: -XX:-UseBASE64Intrinsics, which turns off the Base64 intrinsics, and -XX:UseAVX=2, which turns off all AVX-512 code paths and can slow other work. The Hacker News says no vendor has confirmed either option for this flaw, so test before relying on them.

Don't forget bundled runtimes. Many applications ship their own JRE inside the install directory or container image instead of using the system Java. Patching the operating system's Java package will not fix a Jira server or Jenkins agent that carries its own runtime. Inventory every JVM: check container base images, vendor appliances and anything installed with a bundled runtime. Then follow each vendor's guidance for updating it.

Triage the fleet: Rank systems by three questions: Is it x86 with AVX-512? Does it run an affected Java line? Does it accept untrusted input from the network? Hosts that answer yes to all three go first. ARM64 hosts are not affected by this specific bug, per CyberScoop.

Watch for crashes: Because researchers say a crash is the likely outcome of failed exploitation, unexpected JVM crashes on internet-facing Java services deserve a closer look before you write them off as routine instability. Keep crash logs and error reports, since they may be your earliest sign of probing.

Compensating controls while you patch: Make sure internal tools such as Jenkins and Jira are not exposed to the open internet without a VPN or identity-aware proxy in front. That shrinks the unauthenticated attack surface for this bug and for whatever comes next.

For CMMC and federal contractors: CISA has issued a patch notice, but as of CyberScoop's reporting the flaw is not in KEV, so no binding federal remediation deadline is attached. Record it in your vulnerability-management process and patch on your critical or high timeline anyway.

The bottom line

Java4Shell may never become another Log4Shell. The exploit conditions are narrower: specific Java versions on specific x86 hardware, and Oracle itself rates exploitation as difficult. But it sits in code that much Java server software runs on untrusted input, a research group says it achieved pre-login code execution in identity and DevOps platforms, and fixed builds are already available. That makes it an easy call: patch the November Java builds now, starting with whatever sits in front of your logins and your build pipeline.

Sources