⌁AI·CYBER·BRIEF▌
Vulnerabilities7 min read

AI-found Java4Shell flaw hits Java 17+ on modern x86 servers — patch Keycloak, Jira, Jenkins first

A Codex-assisted researcher found CVE-2026-114938 in Java's Base64 decoder for under $200. Fixes shipped November 17; here's who is exposed and how to mitigate.

A researcher at security firm Calif used OpenAI's Codex coding agent to find Java4Shell (CVE-2026-114938), a memory-corruption flaw in the Java runtime's built-in Base64 decoder that the researchers turned into pre-login remote code execution against default installs of Keycloak, Jira, Confluence, Jenkins, Artifactory and PingFederate. Oracle and the OpenJDK distributions shipped fixes on November 17, 2026. Only x86-64 servers whose processors support AVX-512 VBMI are exposed. If you run Java 17 or newer on recent Intel Xeon or AMD EPYC hardware, update the Java runtime under every service (bundled copies and containers included). Until you can, switch the vulnerable code path off with a JVM flag.

What happened, in plain English

Base64 is a way of writing binary data as plain text. It is everywhere: login tokens, email attachments, API requests, certificates. Java, the platform behind a large share of enterprise software, has a standard Base64 decoder that almost every Java application uses, often on data that arrives from the internet before anyone has logged in.

To make decoding fast, the Java Virtual Machine (JVM, the engine that runs Java programs) swaps the normal decoder for a hand-tuned shortcut, called an "intrinsic", on processors that support AVX-512 VBMI. That is a set of advanced instructions found in recent server chips. According to Calif's write-up, this shortcut gets its arithmetic wrong for certain malformed inputs and writes past the end of the memory it was given.

Think of a mailroom clerk who is told to file a letter in box 40 of 40 and keeps going into the neighbour's cabinet. In a memory-safe language like Java that should never happen. Here it does, because the shortcut is low-level machine code rather than Java.

The impact depends on the application. Calif says it built working attacks that run code before login on Keycloak, Atlassian Jira and Confluence, JFrog Artifactory, Jenkins and PingFederate in their default configurations. It also built post-login attacks on Kafka, Solr, Elasticsearch, Gerrit and Spinnaker. Calif stresses there is no single universal exploit: each application needs its own work.

The AI angle is the method. Calif says researcher Quang Luong pointed Codex at the JVM's optimized code, and it found the bug within an hour on October 24, 2026, for under $200 in total token cost.

Are you affected? What to do now

You are exposed if all of these are true:

  • The service runs on Java 17.0.6 or newer (17, 21, 25 or 27 lines), from any vendor: Oracle, Temurin, Corretto, Microsoft, Zulu, Liberica, SapMachine and Red Hat builds are all listed by Calif.
  • The server is x86-64 with AVX-512 VBMI. Calif names Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD EPYC/Ryzen Zen 4 or newer. Cloud instances on these chips count. ARM servers (e.g. AWS Graviton) are not affected.
  • The Base64 intrinsic has not been disabled (it is on by default).

Java 11 and standard Java 8 are not listed as affected by the Canadian Centre for Cyber Security or, according to Tenable, by the OpenJDK advisory. Oracle's Java 8 Enterprise Performance Pack (8u511b33-perf) is affected.

Fixed versions (Oracle / OpenJDK numbering, per the Cyber Centre): 17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1, and 8u513-perf. GraalVM for JDK 17.0.21.1 and 21.0.13.1, and GraalVM Enterprise 21.3.20.1. Other vendors may use their own build numbers, so check their release notes.

Checklist, in order:

  • Inventory Java. List every Java runtime, including the copies bundled inside products (Jira, Jenkins, Keycloak, Artifactory, security tools) and inside container images. Check the version with java -version, or jcmd <PID> VM.version for a running process.
  • Check the CPU. On Linux, grep -om1 -E 'avx512[_]?vbmi' /proc/cpuinfo returns a match on affected hardware (Datadog's check).
  • Patch internet-facing services first: identity providers (Keycloak, PingFederate), Jira/Confluence, Jenkins, artifact repositories.
  • Rebuild containers. Keycloak, for example, says its 26.8 images were rebuilt on Red Hat OpenJDK 21.0.13.1 (tags latest, 26.8, 26.8.2, 26.8.2-3), and tags 26.8.2-1 and 26.8.2-2 bundle an affected runtime. Keycloak warns not to judge by the application version alone.
  • If you can't patch today, mitigate. Start the JVM with -XX:-UseBASE64Intrinsics (the option Keycloak describes as Red Hat's preferred one). The Cyber Centre suggests -XX:UseAVX=2 as an alternative. Datadog notes that this option is x86-64 only and that an ARM JVM will refuse to start with it. Confirm the flag took effect with jcmd <PID> VM.flags -all.
  • Ask your vendors. For appliances and SaaS-adjacent products that bundle Java, ask whether they run on affected hardware and when they ship a fixed runtime.
  • Watch logs. Look for crashes of Java services on affected hosts. A JVM crash after unusual Base64 input can be a failed exploit attempt. If you use Cloudflare, it released six WAF rules on November 18 (default action: Log, paid plans), which it calls "a compensating control, not a fix."

Exploitation status: as of November 19, the flaw was not in CISA's Known Exploited Vulnerabilities catalog, per Tenable. GreyNoise reported exploitation attempts but no confirmed successful compromise against its sensors as of November 18, also per Tenable. The Canadian Cyber Centre says it is "not aware of active exploitation at this time, but that can change quickly." Calif published only a crash-reproducer and is holding back its application exploits. Others, including VulnCheck, say they have working exploits.

The expert view

Mechanically, this is a classic length-calculation bug in hand-written vector code. When the input ends in a malformed padded fragment, the computed output length goes negative. That value feeds a masked AVX-512 store, which then writes beyond the destination array. The Cyber Centre classifies it as a heap-based buffer overflow (CWE-122). Datadog notes that the basic and URL decoders reach the vulnerable path, while the MIME decoder does not, and that denial of service is the more readily achievable outcome. Turning a 62-byte overwrite into code execution is application-specific work. Calif's point is that AI agents did most of that per-application work too.

Three things stand out.

The bug was hiding in plain sight. Calif's timeline shows a crash report on GitHub on June 9, 2026, a root-cause analysis on September 8, and a proposed fix on September 9 that was closed without merging. Nobody treated a decoder crash as a security bug. Calif says Codex rediscovered it independently and that the agent transcripts show no lookup of the earlier reports. Whether or not that holds, the lesson is that crash reports in core runtimes are now cheap to triage at scale, for attackers as well as defenders.

"Memory-safe language" is a property of your code, not your stack. Java applications inherit the JVM's native code, its compiler intrinsics and its CPU-specific fast paths. The same CPU-dependence means two identical deployments can differ in exposure purely because of the hardware generation they landed on, which complicates scanning and inventory.

The economics have shifted. Under $200 to find, about two more hours for a first working exploit, and agents for the per-target chains: that is the pattern Google described on November 17 and the reason Project Zero is shortening its disclosure deadline. The fix shipped in the same November 17 Oracle out-of-cycle update as the MySQL flaw we covered. The window between patch and exploit should be assumed to be days, not weeks.

Still unclear:

  • Severity. Calif and Datadog cite CVSS 4.0 9.2 (critical). Tenable and BleepingComputer report Oracle's CVSS 3.1 score as 8.1 with high attack complexity. We could not read Oracle's score directly.
  • Credit names. Oracle credits "Quang Luong of Calif.io" and "Albert Yang". Calif credits the original GitHub reporters as sundarms and albertnetymk.
  • Bundled runtimes. Exact fixed build numbers for several vendor distributions were not published on the pages we read.
  • Timing. Calif said on X that details had leaked before a fix, according to Tenable. That may be why the disclosure was rushed.

Official sources

Get the daily brief

AI + security signal by email: headlines, a two-line summary, a link. No noise, no spam.

How often