+− THE DAILY DIFFdev & AI news
SHIP IT

Java4Shell (CVE-2026-114938): a 65-character string and OpenJDK's Base64 decoder

Calif reports that a 65-character Base64 string (64 letter A's and an equals sign) makes OpenJDK's AVX-512 Base64 decoder on x86-64 write 62 bytes past the end of a 50-byte array, before Java rejects the input with an exception.

Want to see our stories first on Google? Add us to your preferred sources — it helps us grow.

Calif reports that a 65-character Base64 string (64 letter A's and an equals sign) makes OpenJDK's AVX-512 Base64 decoder on x86-64 write 62 bytes past the end of a 50-byte array, before Java rejects the input with an exception. OpenJDK and Oracle published fixes on November 17 (CVSS 3.1 8.1; Calif scores it 9.2 on CVSS 4.0). Calif says it turned the overwrite into code execution in eleven products and withholds the recipes. CISA urged patching; the CVE was not on its Known Exploited Vulnerabilities list when checked on November 18.

Read the written edition (English) ↗

What this video covers

  • The bug sits in a hand-written x86-64 assembly routine that HotSpot's C2 compiler swaps in for Base64 decoding; for the 65-character input Java allocates a 50-byte array and the routine writes a full 64-byte block starting at offset 48, so 62 bytes land past the end. Calif's account, reproduced here as theirs; we ran nothing.
  • The length arithmetic underflows to minus one for a padded tail, the mask instruction treats that as all ones, and Java throws IllegalArgumentException only after the write has happened.
  • Scope: OpenJDK 17.0.6 and newer (OpenJDK's matrix marks 17, 21, 25 and 27), on x86-64 CPUs with AVX-512 VBMI (Intel Ice Lake and later, AMD Zen 4 and later). ARM64 is not affected. Standard Java 8 and 11 are not on this CVE; Oracle also lists its Java 8 performance build.
  • Calif says it reached code execution in eleven products in default setups (Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate without a login; Kafka, Solr, Elasticsearch, Gerrit and Spinnaker after one). Those are Calif's own tests; the product recipes are withheld and Calif says there is no universal exploit.
  • Confirmed by vendors: OpenJDK and Oracle list CVE-2026-114938 (CVSS 3.1 8.1); Oracle JDK 17.0.21.1 was released on November 17; Corretto has fixed builds; Temurin is releasing as testing completes; Red Hat is expediting fixes. The public OpenJDK November security branch contains a four line change titled Base64 optimization improvements.
  • Mitigation: update to a fixed build, or start the JVM with -XX:-UseBASE64Intrinsics on any chip (-XX:UseAVX=2 works on x86-64 only and prevents ARM JVMs from starting). On Linux, Calif's check prints four CPU flags; all four (avx512f, avx512bw, avx512vbmi, bmi2) must appear for a machine to be in scope. Rebuild container images.
  • CISA urged immediate patching on November 17; the CVE was not in its Known Exploited Vulnerabilities catalog (version 2026.11.17) when checked on November 18. The security firm Hadrian says it crashed login checks and reached code execution in a lab; that is its own claim. Verdict: SHIP IT.
  • Visual credits: excerpts and a short recording of calif.io/java4shell are credited to Calif.io; thumbnail portrait of James Gosling, Java's creator, by Peter Campbell, CC BY-SA 4.0 via Wikimedia Commons, cropped and adapted by The Daily Diff, he did not author, approve or react to this research; a two-second silent quotation from The Naked Gun (1988) via GIPHY, the film's owners retain their rights.

Questions and answers

What is Java4Shell, and why does Java's error arrive too late?

Java4Shell is CVE-2026-114938, a bug in OpenJDK's x86-64 Base64 decoder routine. Calif reports that a 65-character input makes it write 64 bytes into a 50-byte array, so 62 bytes land past the end. Java then throws IllegalArgumentException, but the write has already happened, which is why neighboring objects may be corrupted.

Watch this part 0:26 · Calif.io (Quang Luong) · OpenJDK

Which Java versions, CPUs and servers are affected by CVE-2026-114938?

OpenJDK's advisory marks Java 17, 21, 25 and 27 builds. Calif and Red Hat say exposure needs an x86-64 CPU with AVX-512 VBMI, such as Intel Ice Lake or later and AMD Zen 4 or later. ARM64 is not affected, and standard Java 8 and 11 are not on this CVE. The compiler must also have warmed up the decoder.

Watch this part 3:29 · Calif.io (Quang Luong) · OpenJDK · Red Hat

What is confirmed about Java4Shell, and what is only Calif's claim?

OpenJDK and Oracle confirm the CVE at CVSS 3.1 8.1, Oracle JDK 17.0.21.1 shipped on November 17, and CISA urged patching. Calif's results are its own tests: code execution in eleven products, a 9.2 score on CVSS 4.0, and an AI effort of about three hours. Calif withholds the product recipes, and the CVE was not on CISA's known exploited list when checked.

Watch this part 3:01 · Calif.io (Quang Luong) · OpenJDK · CISA

How do I check my servers and mitigate Java4Shell today?

Update to a fixed build from your vendor. Until then start the JVM with -XX:-UseBASE64Intrinsics, which works on every chip, per Calif and Adoptium. On Linux, Calif's grep line prints four CPU flags, avx512f, avx512bw, avx512vbmi and bmi2; all four must appear for a machine to be in scope. Rebuild container images, since the Java inside an old image stays old.

Watch this part 4:00 · Calif.io (Quang Luong) · Adoptium · Red Hat

What was the verdict on Java4Shell?

The Daily Diff's verdict was SHIP IT. Ship the patch because it is small, public and already built by the main vendors, and the one startup flag works everywhere. Calif's code execution claims stay Calif's until someone else shows them, so update without panic.

Watch this part 4:43 · Calif.io (Quang Luong) · OpenJDK · Oracle

Transcript

Can a memory-safe language really write out of bounds?

0:00 Java is the memory-safe language. Sixty-four letter A's and an equals sign just made it write sixty-two bytes past the end of an array. In this video, how does a memory-safe language do that? Why does the error message arrive too late? And what do you change on your servers today? Also, which of your servers can even run the broken code? One line of shell answers most of it, and I'll show you before the verdict.

0:22 It's Wednesday, November eighteenth, and this is The Daily Diff.

What is Java4Shell?

0:26 Java4Shell is the nickname for a bug in OpenJDK, the open source Java runtime that Corretto, Temurin and Oracle's builds all come from. Calif, a security research group, found it in the code that decodes Base sixty-four, the text format behind logins and file uploads. OpenJDK and Oracle published fixes on Tuesday. Here's the short answer. Java is memory safe, but the runtime under it has C, C plus plus and hand-written assembly, and that part isn't.

0:53 Once a program decodes Base sixty-four often enough, the compiler swaps in an assembly routine, and this one miscounts the padded end. It needs an Intel or AMD chip with A V X five twelve, and one startup flag switches it off.

Who found it, and how fast?

1:06 Quang Luong at Calif told the Codex coding agent to audit the compiler, the crypto and the common libraries. Most of what came back was junk, he says, then it flagged the Base sixty-four decoder. Calif says that took one hour to find and two more to build an exploit, for under two hundred dollars of tokens. Same team as PHP2Shell, which I covered last month.

How do 64 A's overflow a 50-byte array?

1:27 Base sixty-four packs three bytes into four characters, so sixty-four A's decode to forty-eight bytes. Add the equals sign for padding, and Java reserves a fifty byte array for the answer. The assembly finishes that tail with one sixty-four byte store and a mask that says which bytes to keep. Here the length math reaches zero, the padding takes one more, and the length becomes minus one. The mask reads that as a huge number, so every bit stays on.

1:53 The store starts near the end of the array and writes a full sixty-four bytes. Sixty-two of them land past the end, on whatever object the allocator put next door. Then the decoder notices the input is malformed and throws an exception.

Why does the error message arrive too late?

2:06 The exception is correct. It's also late. The write already happened, so Java rejects the request politely while the damage sits in the heap.

How does a spill turn into code execution?

2:14 The spill is mostly a fixed pattern, so Calif aims it at the length field of a neighboring array. A bigger length makes that array accept bytes the attacker picks. Do it again to a second array that gets copied into the response, and heap memory comes back out. Calif calls that Heartbleed style. From there, Calif says it got code execution in eleven popular Java products. Six, including Keycloak, Jira and Jenkins, needed no login. Five more, including Kafka and Elasticsearch, did.

2:42 Those are Calif's own tests, in default setups. It's holding the recipes back until people patch. Calif is candid about the limits. There's no universal exploit. Every product needs its own recipe, timing is sensitive to network jitter, and their fix for jitter was renting a server in the same cloud region as the target. Physics, plus money.

What is confirmed, and what is only Calif's claim?

3:01 Outside Calif, here's what's confirmed. OpenJDK and Oracle list the bug at eight point one, where Calif says nine point two on a newer scale. A four line change in the public November security branch is titled Base sixty-four optimization improvements. Nothing in that title says security. Cisa urged patching, and when I checked, the bug was not on its known exploited list. The security firm Hadrian says it crashed login checks in a lab,

3:27 and ran code in a lab service.

Which of your servers can run the broken code?

3:29 Now your servers. Four things have to line up. A chip with A V X five twelve V B M I, so Intel from Ice Lake on or AMD from Zen four on. ARM machines like Graviton are out. Java seventeen or newer, and Oracle's Java eight performance build counts too. A decoder the compiler has warmed up. And input an attacker can reach.

3:49 Here's the check I promised. This line prints four C P U flags, and a machine needs all four before the routine can run there. If one is missing, you're out of scope for this bug, though you'd still update.

What do you change today?

4:00 The fix is a new build for each Java line. Oracle shipped one Tuesday, Corretto has its builds, and Temurin is rolling out. Until you update, start Java with minus X X colon minus Use Base sixty-four Intrinsics, which works on every chip. The other flag, Use A V X equals two, stops ARM machines from starting. Rebuild container images too. And if you'd rather read this than hear me say it, the diff lands in your inbox every morning, free at the daily diff dot dev, link below.

4:29 Calif's lesson is to delete things. The fast decoder helps programs that decode enormous amounts of Base sixty-four, and everyone else carried it too. They call the cleanup vibe-uncoding, a real phrase on a real security page. So today's verdict is SHIP IT.

What is the verdict on Java4Shell?

4:45 I'd ship the patch because it's small, public and already built by the main vendors, and the one flag works everywhere. Calif's code execution claims stay Calif's until someone else shows them, so don't panic, just update. Subscribe, hit the bell, and tell me in the comments if you'd have stamped it differently. And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

Sources

  1. Java4Shell: Write once, RCE everywhereCalif.io (Quang Luong)
  2. OpenJDK Vulnerability Advisory: 2026/11/17OpenJDK
  3. Oracle Critical Security Patch Update Advisory - November 2026Oracle
  4. Merge CSPU26_11 into master (PR 33936)OpenJDK / GitHub
  5. Critical Java Vulnerability Requires Immediate PatchingCISA
  6. Known Exploited Vulnerabilities catalog (checked November 18, 2026)CISA
  7. RHSB-2026-018: java4shellRed Hat
  8. Eclipse Temurin guidance for CVE-2026-114938Adoptium
  9. Corretto 25.0.5.1.2 changelogAmazon Corretto
  10. CVE-2026-114938Amazon Linux Security Center
  11. JDK 17.0.21.1 Release NotesOracle
  12. Java4Shell (CVE-2026-114938): pre-auth memory corruption in OpenJDKHadrian
  13. Java4Shell: RCE in OpenJDK when processing Base64Hacker News

Niko from Axrisi

Nikoloz Turazashvili · Vexrail · Axrisi

LinkedIn · dev.to · GitHub · X · Substack

Related videos