WED, NOVEMBER 18, 2026
Independent · In‑Depth · Practitioner‑Tested
✎ Code Tools

An AI Found a Critical Java Flaw in an Hour. Here Is What You Actually Need to Patch.

CVE-2026-114938 is a memory-corruption bug in the Base64 decoder of the x86-64 HotSpot JVM. It only reaches machines with AVX-512 VBMI, ARM64 is unaffected, and Oracle and the researchers who found it disagree on severity: 8.1 against 9.2.

By AIToolsRecap November 18, 2026 6 min read 57 views
Home › Articles › Code Tools › OpenAI Codex › Java4Shell (CVE-2026-114938): What to Patch, an...
An AI Found a Critical Java Flaw in an Hour. Here Is What You Actually Need to Patch.
Java4Shell, in brief
● What it is CVE-2026-114938, nicknamed Java4Shell. Memory corruption in the Base64 decoder of the x86-64 HotSpot JVM.
● Who is actually exposed Only x86-64 chips with AVX-512 VBMI - certain Intel Xeon families, AMD Zen 4 and newer. ARM64 is not affected.
● Patch to Java 17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1, or 8u513-perf for the Enterprise Performance Pack.
● Severity Disputed. Oracle scores it 8.1 on CVSS 3.1 with high attack complexity. Calif, who found it, scores it 9.2 on CVSS 4.0.
● Exploited yet No public exploitation reported as of 17 November, and it is not in CISA's Known Exploited Vulnerabilities catalog.

OpenJDK published a security advisory on Tuesday 17 November and Oracle shipped patched builds the same afternoon, inside a November update covering 652 flaws in total. CISA has urged organisations to patch immediately.

The story is being carried in places as a critical bug in Java affecting most large companies. Two things are worth correcting before you decide how fast to move: the exposure is narrower than all of Java, and the two organisations closest to the bug do not agree on how severe it is.

What to patch

Release lineFixed version
Java 1717.0.21.1
Java 2121.0.13.1
Java 2525.0.5.1
Java 2727.0.1.1
Java 8 Enterprise Performance Pack8u513-perf
Standard Java 8 and Java 11Not listed as affected in the OpenJDK advisory
GraalVMRefer to Oracle's advisory - no version published in the summaries
Corretto, Temurin and other buildsCheck each distributor; they patch on their own schedule

The OpenJDK Vulnerability Group lists the 17, 21, 25 and 27 release lines as affected. Calif puts the floor at OpenJDK 17.0.6 and newer. Those two statements have not been reconciled publicly, so if you are on a 17.0.x build below 17.0.6 do not assume you are clear on Calif's say-so alone.

Oracle names these builds as affected: 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, plus GraalVM for JDK 17.0.21.0.3 and 21.0.13.0.3 and GraalVM Enterprise Edition 21.3.20.0.3 - and adds that earlier versions are likely affected too.

Check whether you are actually exposed

This is the part the breaking-news framing drops. The vulnerable code is a hand-written optimisation the JVM uses to speed up Base64 decoding, and it only runs on x86-64 processors that support AVX-512 VBMI. On anything else the fast path is not taken and the bug is not reachable.

On Linux, check whether your CPU reports all four of avx512f, avx512bw, avx512vbmi and bmi2. In practice that means certain Intel Xeon families and AMD Zen 4 or newer. ARM64 is not affected, which rules out Graviton fleets and Apple Silicon development machines in one line.

That is a real reduction in scope, and it is the difference between a weekend and a quarter of work. It is not a reason to skip patching - it is a reason to know which hosts to do first.

Two scores, and why they differ

SourceScoreStandard
Oracle8.1, high attack complexityCVSS 3.1
Calif Research (the finder)9.2CVSS 4.0

These are not the same scale. CVSS 4.0 changed how attack requirements and exploit maturity are weighted, so a 9.2 under 4.0 and an 8.1 under 3.1 are not directly comparable, and neither party is wrong. What matters operationally is Oracle's high attack complexity rating: it means the conditions needed to reach the bug are not trivially arranged, which is consistent with the AVX-512 requirement.

If your vulnerability management tooling ingests one score and not the other, you may see this ticket ranked very differently depending on which feed it came from. Worth knowing before someone escalates or deprioritises it on a number alone.

If you cannot patch today

Two JVM flags disable the vulnerable path. Neither appears in Oracle's or the OpenJDK advisory, so treat them as community mitigations and test before rolling out:

-XX:-UseBASE64Intrinsics turns off the Base64 fast path specifically. -XX:UseAVX=2 disables AVX-512 more broadly - and note that this one is x86-64 only, because Java on ARM will not start with it set.

Both trade some Base64 throughput for safety. On a service that decodes a lot of Base64 - which is most things handling JWTs, attachments or image payloads - measure it rather than assuming the cost is negligible.

The part that is actually new

Quang Luong at Calif Research found this while auditing OpenJDK with OpenAI's Codex. It took about an hour and under $200 in compute tokens. Calif says AI agents produced most of its exploit chains for the work.

A heap overflow in the Base64 decoder of the most widely deployed runtime in enterprise software, in code that has been read by an enormous number of competent people, found in an hour for the price of a cheap monitor.

That is the same pattern as GPT-6 Cyber surfacing 81 unknown high-severity Firefox flaws and as Fortinet patching AI-discovered bugs before publishing CVE details. The interesting question is no longer whether these tools find real vulnerabilities. It is what happens to disclosure timelines when auditing a major codebase costs a few hundred dollars and anyone can do it.

Researcher samczsun separately demonstrated that Java programs can be compromised by connecting to a malicious server, which is what moves this from a parsing bug to something worth the CISA alert.

FAQ

Does Java4Shell affect every Java application?
No. The vulnerable code only runs on x86-64 processors supporting AVX-512 VBMI - certain Intel Xeon families and AMD Zen 4 or newer. ARM64 is not affected. The OpenJDK advisory lists the 17, 21, 25 and 27 release lines; standard Java 8 and Java 11 are not listed as affected.
What version do I need to be on?
17.0.21.1, 21.0.13.1, 25.0.5.1 or 27.0.1.1 depending on your release line, or 8u513-perf for the Java 8 Enterprise Performance Pack. If you run a distribution other than Oracle's - Corretto, Temurin and so on - check that vendor, because they patch on their own schedule.
Is it being exploited?
No public exploitation had been reported as of 17 November 2026, and the CVE is not in CISA's Known Exploited Vulnerabilities catalog. That can change quickly once a flaw of this profile is public.
Why do I see two different severity scores?
Oracle rates it 8.1 using CVSS 3.1 and flags high attack complexity. Calif Research, who discovered it, rates it 9.2 using CVSS 4.0. Those are different versions of the scoring standard and are not directly comparable.
Can I mitigate without patching?
Two JVM flags disable the vulnerable path: -XX:-UseBASE64Intrinsics, or -XX:UseAVX=2 to turn off AVX-512 more broadly. Neither is in Oracle's or the OpenJDK advisory, both cost some Base64 throughput, and the AVX one is x86-64 only - Java on ARM will not start with it.
How was it found?
Quang Luong at Calif Research found it in about an hour while auditing OpenJDK with OpenAI's Codex, for under $200 in compute tokens.
Tags
AI NewsOpenAICoding AI2026

Spot an inaccuracy?

We verify facts before publishing and correct errors promptly. If something in this article is wrong or outdated, let us know.

Report an error →