Safety & Security

Calif Used AI to Find Java4Shell, a 9.2-Severity OpenJDK Flaw, in One Hour

Calif used AI to find Java4Shell (CVE-2026-114938), a 9.2-severity OpenJDK Base64 flaw enabling pre-auth RCE. Fixes shipped the same day for the entire Java ecosystem.

Calif Uses AI to Find Java4Shell as OpenJDK and Oracle Ship Fixes
Calif Uses AI to Find Java4Shell as OpenJDK and Oracle Ship FixesWinston Chen / unsplash
By Elena Vasquez5 min read

Why it matters

  • Calif disclosed Java4Shell (CVE-2026-114938) on November 17, 2026, rated 9.2 under CVSS 4.0; OpenJDK and Oracle shipped fixes the same day.
  • AI-assisted discovery took one hour, the first working exploit two more hours, with token costs below $200.
  • The flaw enables pre-authentication RCE in default configurations of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate.
  • Affected runtimes include OpenJDK 17.0.6+, Amazon Corretto, Eclipse Temurin, Microsoft Build of OpenJDK, Azul Zulu and Oracle Java SE, on x86-64 CPUs with AVX-512 VBMI only.
  • Oracle's November advisory contains 11 new Java SE patches and credits Quang Luong of Calif.io and Albert Yang for the CVE.

Calif disclosed a critical OpenJDK vulnerability on November 17, 2026, rated 9.2 under CVSS 4.0, that lets an unauthenticated attacker corrupt memory and execute code remotely — and the security firm found it with AI in a single hour, for under $200 in token costs. OpenJDK and Oracle shipped fixes the same day.

The flaw, tracked as CVE-2026-114938 and nicknamed Java4Shell, sits in hand-written x86-64 assembly used by OpenJDK HotSpot to accelerate Base64 decoding with AVX-512 VBMI instructions. Its location matters because it undercuts a familiar Java promise: applications benefit from managed memory and memory-access checks, but this part of the runtime is written in assembly, where those protections do not eliminate memory-corruption risk.

Who found it, and how?

Calif reports that researcher Quang Luong directed Codex, an AI coding model, to examine the JIT compiler, intrinsics, widely accessible APIs and cryptography. The model surfaced the Base64 finding, with discovery taking one hour, the first working exploit another two hours, and total token costs below $200. Luong chose the targets and supplied exploitation expertise. Calif built pipelines and a heap tracer and visualizer to help AI agents develop later exploit chains.

The stakes are significant for the Java ecosystem, which runs a large share of enterprise infrastructure. The bug affects OpenJDK 17.0.6 and later, plus every distribution built from OpenJDK: Amazon Corretto, Eclipse Temurin, Microsoft Build of OpenJDK, Azul Zulu, Red Hat OpenJDK, Oracle Java SE and Oracle GraalVM for JDK. OpenJDK marks Java 17, 21, 25 and 27 as affected and classifies the issue under hotspot/compiler.

What does the bug actually do?

The optimized routine mishandles padding in malformed Base64 input. A 65-character string containing 64 A characters and an equals sign produces a 50-byte output allocation. An invalid negative output length then enables a 64-byte write that sends 62 bytes beyond that allocation. Java eventually rejects the input, but neighboring heap objects may already be corrupted by then.

For OpenJDK 17, the report gives a different trigger — 128 A characters plus an equals sign — because its threshold for reaching the affected code path is higher.

Calif's exploit chain goes further than a fixed overwrite. Calif reports that its chain corrupts a buffer length to upgrade the fixed overwrite into chosen-value writes, then uses another corrupted array to disclose heap contents such as private keys and tokens.

Which products are vulnerable?

The demonstrated impact list is long. Calif showed pre-authentication RCE in default configurations of:

  • Keycloak
  • Atlassian Jira
  • Atlassian Confluence
  • JFrog Artifactory
  • Jenkins
  • PingFederate

The researchers also demonstrated post-authentication RCE in Apache Kafka Broker, Apache Solr, Elasticsearch, Gerrit and Netflix Spinnaker, plus RCE in sample Apache Struts and Spring Boot applications. Calif notes the published product lists are non-exhaustive.

A vulnerable Java process does not even need to expose a service of its own. Researcher samczsun separately shared a client-side exploit with Calif showing that a Java program opening a TLS connection to a malicious server can be compromised.

Why isn't this a universal exploit kit?

Hardware and targeting constraints blunt the blast radius. The vulnerable code requires an x86-64 processor with AVX-512 VBMI support. ARM64 processors are not affected. Listed processors that meet the hardware requirement include Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, plus AMD EPYC and Ryzen processors based on Zen 4 or newer.

Severity ratings differ by scoring system. Calif's 9.2 critical score uses CVSS 4.0. The OpenJDK advisory and Oracle's November update assign 8.1 under CVSS 3.1. Oracle assesses the flaw as remotely exploitable over a network without authentication, with high attack complexity and high impacts on confidentiality, integrity and availability.

Exploitation also demands work per target. Calif says each target needs a bespoke profile containing a decoder route, a live writer, a reader or another feedback path, a suitable heap layout and an application action. The firm's remote-grooming approach is sensitive to round-trip time and jitter, and incorrect object placement can crash the JVM — though retries are often possible.

Buffering complete request bodies can defeat some grooming profiles, but not all. The Elasticsearch chain works from one complete request, and configured WebSocket tunnels change the buffering conditions.

What are the fixes and mitigations?

OpenJDK's November 17, 2026 advisory recommends upgrading as soon as possible. Oracle's JDK 17.0.21.1 release notes carry the same release date and link the security fixes to Oracle's November advisory, which contains 11 new Java SE patches and credits Quang Luong of Calif.io and Albert Yang for CVE-2026-114938.

The mainline OpenJDK change is small but surgical: a four-line guard that sends a padded Base64 tail shorter than one complete block to the Java handler rather than the affected optimized routine.

Until updates are applied, Calif lists two temporary mitigations:

  • -XX:-UseBASE64Intrinsics — disables hardware-accelerated Base64 encoding and decoding
  • -XX:UseAVX=2 — disables AVX-512, but only on x86-64 hosts; OpenJDK on ARM does not recognize it and will fail to start

Calif published the vulnerability mechanics and a bounded standalone trigger, but is withholding application-specific exploit chains while users upgrade. That limits the immediately available material to the underlying flaw and general exploitation technique rather than ready-made attacks against individual products.

Does the story end with a patch?

Calif's broader argument reaches past this single bug: AI can also help maintainers identify and remove unnecessary features and optimizations. Java4Shell gives that position a concrete case. The Base64 optimization was intended for applications decoding enormous amounts of data, yet it shipped to every affected runtime — exposing applications that gained little from it.

The same AI-assisted workflow that found the bug in one hour is thus being presented as a way to reduce attack surface, not only to discover and weaponize flaws. For runtime maintainers weighing which optimizations earn their risk, Java4Shell is now the reference case.

Source: unite.ai

Share this article:

More from Elena Vasquez

Elena Vasquez

Show full bio

Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.

708 articles

Related articles

  1. OpenAI's cyber model found 948 Linux root exploits for $1.2 million
  2. AI found a Google zero-day for $4.24 in 13 minutes in 2026
  3. Six Fake SQLite CVEs and a 10x Surge Expose a Broken Pipeline
  4. 85% of Security Leaders Report Patching-Delay Incidents, Azul Survey Finds
  5. GPT-6 Cyber found 81 unknown Firefox flaws and 948 kernel exploits

« Previous articleNext article »