Inter-office memo

To
Everyone running Java 17 or later
From
Mo, Chief Executive
Date
November 18, 2026, 12:14 PM CST
Re
Java4Shell

Java4Shell: AI found a critical Java flaw. Who’s at risk, and which builds fix it

On November 17, 2026, Oracle and OpenJDK fixed Java4Shell (CVE-2026-114938), a Base64 bug that can let attackers run code without logging in on Java 17.0.6 through 27 on x86-64 chips with AVX-512 VBMI. The fixes are 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1, or your vendor’s own build. Java 8, 11 and ARM64 aren’t affected. Calif found it with Codex in about an hour.

What Oracle and OpenJDK fixed on November 17

At 1:00 p.m. Pacific on November 17, 2026, the OpenJDK project published a security advisory and Oracle released its November Critical Security Patch Update, both fixing CVE-2026-114938. Calif, the security firm that reported it, calls it Java4Shell: “a memory-corruption vulnerability in OpenJDK’s Base64 decoder.”

The bug is in hand-written machine code Java uses to decode Base64 faster on x86-64 processors with AVX-512 VBMI instructions. An input as short as 64 A characters followed by = makes it write 62 bytes past the end of its output. Java then rejects the input, but “by then neighboring objects may already be corrupted,” Calif says, and it turned that into a leak of memory, “as Heartbleed did,” and into remote code execution.

The scores differ. Calif rates it 9.2 out of 10 under CVSS 4.0. Oracle’s CVE record gives 8.1 under CVSS 3.1, calls it a “Difficult to exploit vulnerability,” and says a successful attack “can result in takeover of Oracle Java SE.” Red Hat rates it Critical.

Every Java vendor ships its own fixed build, with its own number. Here is where the main ones stood on November 18:

Java Exposed Oracle and OpenJDK Amazon Corretto Microsoft Eclipse Temurin
8 and 11 No, except Oracle’s 8u511b33-perf 8u513-perf Not affected Not affected Not affected
17 17.0.6 and later 17.0.21.1 17.0.21.1.2 17.0.21.1 17.0.21.1+2, x64 still in testing
21 Every build before the fix 21.0.13.1 21.0.13.1.2 21.0.13.1 21.0.13.1+2, x64 still in testing
25 Every build before the fix 25.0.5.1 25.0.5.1.2 25.0.5.1 25.0.5.1+2, still in testing
27 Every build before the fix 27.0.1.1 27.0.1.1.4 No 27 build 27.0.1.1+4, out for x64 Linux

Corretto’s fixed builds came out on November 17, and Amazon Linux’s packages with them. Microsoft’s followed on November 18. Temurin is “actively releasing binaries” as testing completes; at 17:10 UTC on November 18, its release tracker showed only JDK 27 for x64 Linux published among its x64 builds. Red Hat says it is “expediting the release of fixes” for RHEL 8, 9 and 10 and its builds of OpenJDK 17, 21 and 25, and asks customers to update “as soon as errata are available.” Oracle’s GraalVM for JDK is fixed in 17.0.21.1 and 21.0.13.1.

Who’s at risk

A Java process is exposed only when all three of these hold, by Calif’s and Red Hat’s accounts:

  1. An affected Java. OpenJDK 17.0.6 or later, 21, 25 or 27, before its vendor’s fix. Calif lists Corretto, Temurin, Microsoft Build of OpenJDK, Azul Zulu, BellSoft Liberica, SAP SapMachine, Red Hat OpenJDK, Oracle OpenJDK, Oracle Java SE and GraalVM for JDK.
  2. A processor with AVX-512 VBMI. That’s Intel Xeon from Ice Lake on (Sapphire Rapids, Emerald Rapids, Granite Rapids) and AMD EPYC and Ryzen from Zen 4 on. Red Hat names the cloud machines “most likely to be affected”: AWS C6i, M6i, R6i, C7i, M7i and R7a; Google Cloud C3, C3D, C4, C4D, N4, N4D, M3 and M4; Azure Dsv5, Dsv6, Dasv6, Dasv7 and related D-series. “ARM64 CPUs are not affected,” Calif says, so neither are AWS Graviton servers or Macs with Apple silicon.
  3. Base64 an attacker controls. Red Hat calls that “common in practice: HTTP Basic credentials, SAML artifacts, CSRF tokens, and MIME encoded words are all Base64-encoded and are routinely decoded before authentication completes.” No setting turns the fast path on: it’s chosen “automatically whenever the JVM runs on a CPU reporting AVX-512 VBMI support.”

Calif ran code without logging in on Keycloak, Atlassian Jira and Confluence, JFrog Artifactory, Jenkins and PingFederate, “in their default configs,” and after logging in on the Apache Kafka broker, Apache Solr, Elasticsearch, Gerrit and Spinnaker. It calls the lists “non-exhaustive,” and it also broke into sample apps built on Apache Struts and Spring Boot. A server isn’t the only target. Calif also has a client-side exploit from the researcher samczsun: “A Java program that opens a TLS connection to a malicious server can be compromised, even if it exposes no service of its own.”

On Linux, Red Hat’s checks are:

java -version
grep -o avx512vbmi /proc/cpuinfo | head -1

“If there is no output, the CPU does not support AVX-512 VBMI and is not known to be affected.”

How AI found it

“As with PHP2Shell, Calif used AI to find and weaponize Java4Shell,” the firm writes, a month after PHP2Shell. Its researcher Quang Luong aimed at Java’s JIT compiler, its hand-tuned fast paths, widely accessible APIs and cryptography: “I told Codex to look at those areas like any serious security researcher in 2026.” Most early results were “slops, either unreachable or the requirement of exploitation is not realistic at all,” and Luong “just prompted harder” until Codex reported memory corruption in the Base64 decoder. Calif doesn’t say which model Codex was running.

The numbers are Calif’s: “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.” After that, “most of the later chains are built almost autonomously with AI.” Not everything went smoothly: “the AI would get stuck during heap grooming,” so Calif wrote “a small heap tracer and visualizer to give the agent information about the heap state.” Cloudflare draws the conclusion for defenders: “we expect other attackers to build exploit chains quickly.”

The bug wasn’t wholly unknown. On September 9, Albert Yang opened an OpenJDK pull request titled “AVX-512 Base64 decoder overwrites output buffer on short padded tail,” with the same trigger, 64 As and an =, and a suggested fix. It was closed on November 4 without being merged. Oracle’s advisory credits both Albert Yang and “Quang Luong of Calif.io” for the CVE.

Is it being exploited?

Not that anyone had reported by November 18:

  • The November 17 version of CISA’s Known Exploited Vulnerabilities catalog doesn’t list it.
  • Canada’s Cyber Centre “is not aware of active exploitation at this time, but that can change quickly.”
  • Cloudflare, which runs Java itself, has “seen no indication so far that Java4Shell was exploited on Cloudflare infrastructure.”
  • Calif published only a self-contained test. “The weaponized chains for individual products stay private until everyone has had a chance to upgrade. We will publish them later.”

Others are catching up. Other labs had reproduced crashes and a memory leak but not code execution, The Hacker News reports, and Tenable says Kevin Breen of Immersive posted on X about having a working proof of concept, without releasing it. Since Calif’s product chains aren’t public, code that claims to be one didn’t come from Calif.

What to do

  1. Find every Java you run, not just the system one. As Cloudflare puts it, “Many applications bundle their own Java runtime, and container images ship with a fixed Java Development Kit (JDK). Upgrading the system Java package does not update those copies, and it does not replace a Java virtual machine (JVM) that is already running.”
  2. Install your vendor’s fixed build, then restart. Check what the running process reports, with jcmd <pid> VM.version, rather than the package. Watch the numbers: Oracle lists 17.0.21.0.3 and 21.0.13.0.3 as affected, and the fixes are 17.0.21.1 and 21.0.13.1.
  3. Where there’s no fix yet, turn off the fast path. Start Java with -XX:-UseBASE64Intrinsics and restart it. Red Hat calls this “the preferred option, as it removes the vulnerable code path while leaving other AVX-512 optimizations in place.” -XX:UseAVX=2 also works, but only on x86-64: “JVMs on ARM64 do not recognize it and will fail to start.” Confirm it took with jcmd <pid> VM.flags -all | grep -F UseBASE64Intrinsics, which should print bool UseBASE64Intrinsics = false. Cloudflare warns the flags “may reduce performance for workloads that decode large amounts of Base64.”
  4. Do Keycloak first. Its advisory says the x86-64 images tagged 26.8.2-1 and 26.8.2-2 “bundle an affected OpenJDK 21.0.13 build,” and rebuilt images will follow patched Java builds. Red Hat adds that “a login requirement does not protect an affected instance.” Then move on to everything else that faces the internet, login services first.
  5. Update Java 8 and 11 anyway. Java4Shell isn’t in them, but the same release fixes nine other OpenJDK flaws, and one of them, CVE-2026-114939, rated 7.4, reaches all six Java lines.
  6. Use a firewall rule as a backstop, not a fix. On November 18, Cloudflare added six rules for Pro, Business and Enterprise, “enabled by default with the Log action”; review the matches in Security Events, then set them to Block. It’s “working on a rule for the Cloudflare Free Managed Ruleset.” Microsoft published custom rules for Azure WAF. Neither sees every path: Cloudflare notes Java4Shell “can be reached through paths a WAF does not see, like outbound TLS responses,” and Kafka speaks its own protocol.

Checking your servers with an agent

Cloudflare’s Java4Shell page has a prompt “with an agent that can inspect your infrastructure,” which you can give Claude Code, Codex or another coding agent running where it can reach your servers. It begins:

Assess whether the systems available to you are affected by CVE-2026-114938 (Java4Shell), a reported remote code execution flaw in OpenJDK. Perform a read-only assessment within your authorized access. Do not attempt exploitation, change systems or restart services.

The rest asks the agent to read Oracle’s and each vendor’s advisories, inventory hosts, container images and bundled JDKs, record each running JVM’s version with java -version or jcmd <pid> VM.version, note which services face the internet, and return a table. It tells the agent to “Separate an installed fix from a running fix,” which is the mistake steps 1 and 2 above guard against. Paste the whole prompt from Cloudflare’s page, and keep it read-only: the result is a list for you to act on.

To test one JDK by hand, Calif’s write-up includes a standalone Base64Trigger program that prints REPRODUCED on an affected Java. Calif says “Every write lands inside a real array, so the trigger is safe to run.” It only shows the bug on a machine with AVX-512 VBMI, so a laptop result doesn’t speak for your servers.

What hasn’t been said

  • Whether anyone is exploiting it. No one has reported attacks, and Calif’s product exploits are private for now.
  • When the rest of the fixes ship. Temurin’s x64 builds of 17, 21 and 25, Red Hat’s packages and Keycloak’s rebuilt images had no dates on November 18.
  • What Atlassian, Jenkins and the other named vendors will do. On November 18, neither Atlassian’s advisories nor Jenkins’ listed Java4Shell, so neither has said whether its own packages need new builds.
  • Whether JetBrains IDEs need an update. They run on JetBrains’ own Java, whose latest release came out about 20 hours before the fix. JetBrains hasn’t said.
  • Which model found it, and why September’s pull request didn’t land. Calif names Codex, not a model. The pull request was closed without a fix merged, and Oracle’s advisory doesn’t mention it.

We’ll update this memo when Temurin and Red Hat ship their fixes, or when Calif publishes its exploits.

8.Questions.

Am I affected by Java4Shell?

Only if three things are true at once. The Java is 17.0.6 or later, 21, 25 or 27, from any vendor, without its November fix; it runs on an x86-64 processor with AVX-512 VBMI, such as Intel Xeon Ice Lake or newer or AMD Zen 4 or newer; and it decodes Base64 an attacker can send, as login headers and tokens often are. Standard Java 8 and 11 and ARM64 machines aren’t affected.

Which Java version fixes Java4Shell?

Oracle’s and OpenJDK’s fixes are 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1, released November 17, 2026. Other vendors number theirs differently, such as Amazon Corretto’s 17.0.21.1.2, so check your vendor’s page. Oracle’s 17.0.21.0.3 and 21.0.13.0.3 are still affected.

Was Java4Shell found by AI?

Yes. Calif says its researcher Quang Luong pointed Codex at Java’s JIT compiler and fast paths, and that finding the bug took an hour, the first working exploit two more hours, and under $200 in tokens. The same bug had also been described in an OpenJDK pull request in September that was never merged.

Is Java4Shell being exploited?

No attacks had been reported by November 18. CISA’s list of exploited vulnerabilities doesn’t include it, Canada’s Cyber Centre isn’t aware of any, and Calif is keeping its exploits for specific products private until people have had time to update.

What do I do if I can’t update Java yet?

Start Java with -XX:-UseBASE64Intrinsics and restart it, which turns off the code with the bug. Red Hat calls it the preferred stopgap; the other one, -XX:UseAVX=2, works only on x86-64, because Java on ARM64 won’t start with it. Either way, install the fixed build when your vendor ships it.

Are Jira, Jenkins and Keycloak affected?

When they run on an affected Java and processor, yes. Calif ran code without logging in on Keycloak, Jira, Confluence, JFrog Artifactory, Jenkins and PingFederate in their default setups. Keycloak says its current x86-64 images bundle an affected Java and tells you to use the stopgap until rebuilt images ship.

Enjoy Agentic Industries · Checked against its sources on