The bug is in Java itself

Most Java vulnerabilities live in a library a team can pin, scan and upgrade. This one lives in the runtime. The flaw sits in hand-written assembly that the HotSpot JVM uses to decode Base64 faster on modern x86-64 chips. Feed the decoder a short malformed input and a length calculation underflows; the routine writes past the end of the array it was given, into whatever object sits next to it on the heap.

That placement is why the affected list reads the way it does. Base64 decoding happens constantly, and much of it happens before a user logs in: authentication headers, tokens, file uploads, API messages. Calif says that reach gave it remote code execution, with no credentials, against default installations of Keycloak, Jira, Confluence, JFrog Artifactory, Jenkins and PingFederate. Five more products, including Kafka, Solr and Elasticsearch, were reached after a login, on Calif's testing. The firm calls both lists incomplete, and it is keeping the product-specific exploit chains private until users have had a chance to upgrade.

Two conditions narrow the exposure, and both are checkable today. The vulnerable code only runs on x86-64 processors with AVX-512 VBMI support, broadly Intel Xeon from Ice Lake onward and AMD Zen 4 or newer. ARM64 servers never run it. And the Java build has to be OpenJDK 17.0.6 or later. Older Java 17 builds, standard Java 8 and Java 11 are outside the range, with one exception: Oracle's Java 8 Enterprise Performance Pack received the same optimization and is affected.

Three scores for one bug

Ask how severe Java4Shell is and the answer depends on who scored it. Calif rates it 9.2 under CVSS 4.0. The OpenJDK advisory and Oracle rate it 8.1 under CVSS 3.1, with high attack complexity. Red Hat's preliminary score is 9.8. Amazon scored it 8.8 and splits the impact differently again: a crash for an unauthenticated attacker, code execution for an authenticated low-privilege user in certain environments.

The independent evidence so far sits between those positions. Upwind, testing on an AWS Ice Lake instance, corrupted the neighboring object in 97% of rounds with an exact-size output buffer and crashed the JVM. Blackpoint Cyber crashed a test application with a burst of malformed login requests and reproduced JavaBleed, Calif's memory-read step, in all 30 of its local runs. Neither lab has reproduced code execution. That does not make Calif's claim wrong; the firm has had the most time on this bug, and it is withholding the chains that would settle the question. It does mean the most alarming sentence in the coverage currently rests on one company's testing.

Patch priority should not wait for the argument to resolve. CISA urged organizations to patch immediately on November 17, and no scorer puts this below high severity. But teams that sort their queues by a single CVSS number should know which number they are holding, because in this case the spread is 8.1 to 9.8 for identical code.

The patch map is the real problem this week

Oracle moved fast. Its November Critical Security Patch Update landed on November 17 at 1 p.m. Pacific with 655 fixes, Java4Shell among them, and fixed Oracle builds for Java 17, 21, 25 and 27. Amazon Corretto and Azul Zulu have fixed builds out as well.

The rest of the map is blank spots. As of November 18, Red Hat had not released fixes for Red Hat Enterprise Linux 8, 9 and 10 or for its OpenJDK 17, 21 and 25 builds, per its security bulletin as reported by SQ Magazine. Eclipse Temurin, Microsoft's Build of OpenJDK and BellSoft Liberica had not listed fixed builds yet. That matters more than the score argument, because the fix reaches a server only when its own source ships it: a distribution build, a Linux package, or the JDK bundled inside a product like Jira or Jenkins, each waiting on its own vendor. Dependency scanners will not close this gap on their own: the flaw is in the runtime, not in a library version they track.

Until a fixed build exists for their distribution, teams have a documented stopgap: launch Java with -XX:-UseBASE64Intrinsics, which turns off the accelerated decoder that contains the bug. The other check is the hardware: Calif's write-up lists the four CPU flags (avx512f, avx512bw, avx512vbmi, bmi2) to look for in /proc/cpuinfo, and a host missing any of them does not run the vulnerable code. Those are the useful actions today. The score debate can continue without them.

Found in an hour, for under $200

Calif says researcher Quang Luong pointed OpenAI's Codex at OpenJDK's compiler, intrinsics and cryptography code, and the model surfaced the Base64 flaw after working through false leads. The firm's account puts discovery at about an hour, the first working exploit at two more hours, and the total token cost under $200. Oracle credits Luong and Albert Yang in its patch update.

Read that claim the way this site reads vendor numbers generally: it is the finding firm's own account, and the hour almost certainly excludes the expertise it took to know where to point the model and how to judge what it brought back. Wiz says its own AI research tool produced a proof of concept for the flaw within two hours of looking. And the September detail is the uncomfortable one: a malformed input that triggers this overflow sat in a public, unmerged OpenJDK pull request from September 9, weeks before the patch. Whether anyone reviewing it flagged the security implication at the time has no public answer; SQ Magazine is asking the question in print.

No attacks in the wild had been reported as of November 18, and CISA had not added the flaw to its Known Exploited Vulnerabilities catalog. Calif plans to publish its product exploit chains later, once users have had a chance to upgrade. For users of the distributions still waiting on a fixed build, that makes the vendor's ship date, not their own patching discipline, the date that matters.

Sources

  1. [1] SQ Magazine, "Java4Shell Patched by Oracle as Red Hat Fixes Lag Behind," November 18, 2026 — vendor fix table (Oracle/Amazon/Azul fixed; Red Hat none yet), Upwind and Blackpoint reproduction results, Red Hat preliminary 9.8, Amazon 8.8, September 9 pull-request detail.Read source
  2. [2] Cyber Kendra, "Java4Shell (CVE-2026-114938): Affected Versions, Fix, Workaround," November 18, 2026 — decoder mechanics, affected builds table as of early November 18, Amazon advisory framing, Wiz proof of concept, no-wild-exploitation status.Read source
  3. [3] Calif, Java4Shell technical write-up, November 17, 2026 (primary source for the flaw, affected range, product testing, discovery account)
  4. [4] Oracle, November 2026 Critical Security Patch Update (primary source for the fix, the 655-fix update, credit to Quang Luong and Albert Yang, 8.1 CVSS 3.1)
  5. [5] OpenJDK security advisory (primary source for CVE-2026-114938 at 8.1)
  6. [6] CISA alert / @CISACyber post, November 17, 2026 — urged immediate patching of CVE-2026-114938