Calif Security Research disclosed Java4Shell, tracked as CVE-2026-114938, a memory-corruption vulnerability in OpenJDK’s Base64 decoder. Calif reported that researcher Quang Luong identified it in roughly an hour while auditing OpenJDK with OpenAI Codex, spending under $200 in compute credits.
How the Bug Actually Works
The vulnerable code path lives in hand-written processor instructions built for x86-64 chips with specific AVX-512 capabilities.
According to Calif, the flaw resides in performance-optimized, hand-written processor instructions inside the x86-64 HotSpot JVM. ARM64 systems are not affected by this vulnerability’s code path.
Calif reported that a specially crafted input triggers the problem: 64 “A” characters followed by a single equals sign. That sequence causes the decoder to miscalculate its output length and overwrite adjacent heap memory.
The immediate result is typically a JVM crash. Under carefully arranged heap conditions, that corruption can potentially convert into unauthenticated arbitrary code execution, and Calif compared the possible information-disclosure impact broadly to Heartbleed, while noting the underlying mechanism differs.
Exploitation depends on the processor, JVM build, application behavior, the input path, and memory layout. Not every Java application is automatically remotely exploitable.
Severity Scores Differ by Source
CERT-EU and Calif assigned different CVSS scores, and organizations should consult vendor advisories rather than treat the figures as interchangeable.
CERT-EU scored CVE-2026-114938 at CVSS 8.1. A Broadcom security bulletin reported a CVSS v4.0 score of 9.2, a figure also cited by Calif and several security analyses.
Affected Versions and Downstream Distributions
Multiple Java branches and most major OpenJDK distributions carry the exposure, subject to relevant processor and build conditions.
CERT-EU listed affected branches as Java 17 through 17.0.21, Java 21 through 21.0.13, Java 25 through 25.0.5, and Java 27 through 27.0.1. CERT-EU also identified Oracle Java SE 8 performance build 8u511-perf as affected, with 8u513 as the corresponding fix.
Fixed builds for the other branches are 17.0.21.1, 21.0.13.1, 25.0.5.1, and 27.0.1.1; organizations should verify exact version identifiers with their Java vendor because packaging details differ across distributions.
Seven major downstream distributions carry the exposure according to CERT-EU: Eclipse Temurin, Amazon Corretto, Azul Zulu, Microsoft Build of OpenJDK, BellSoft Liberica, SAP SapMachine, and Red Hat OpenJDK builds. Applications that process attacker-influenced Base64 data may warrant product-specific review; that group includes Keycloak, Jira, Confluence, Jenkins, and PingFederate, with risk varying by deployment configuration.
Exploitation Status
CISA urged immediate patching after researchers demonstrated plausible unauthenticated code execution in widely deployed applications.
At the time of the initial CISA notice, confirmed exploitation had not been reported, and CVE-2026-114938 had not been added to CISA’s Known Exploited Vulnerabilities catalog as of that alert. A subsequent Qualys report indicated threat actors had begun exploiting the vulnerability in attacks; organizations should check current advisory sources for the latest status.
What to Do Right Now
CISA guidance prioritizes internet-facing services and bundled runtimes inside third-party products.
Every bundled Java runtime, including those inside container images, appliances, and third-party products, needs an immediate inventory per CISA guidance. Priority goes to internet-facing and unauthenticated services; affected container images should be rebuilt or redeployed, and JVM processes restarted after any update.
Where immediate patching is not possible, CERT-EU identifies temporary measures including disabling the Base64 intrinsic, disabling AVX-512 where operationally acceptable, restricting unauthenticated access, validating and limiting Base64 input, and monitoring for unexpected JVM crashes or suspicious Java processes.
What the AI Discovery Timeline Signals
A sub-$200 discovery compresses the window between public disclosure and potential adversary weaponization.
An hour of AI-assisted auditing was enough. Calif reported that Luong found the flaw using OpenAI Codex for under $200 in compute credits, and that AI agents contributed substantially to subsequent exploit-chain development.
That timeline suggests defenders may have less margin between disclosure and weaponization than past research cycles allowed, though hardware requirements, runtime versions, application input paths, and heap layout remain decisive constraints. AI accelerates discovery; it does not remove the conditions that determine whether a given deployment is actually reachable.






























