On 17 November 2026, Oracle and OpenJDK fixed a critical flaw in Java’s Base64 decoder, dubbed Java4Shell (CVE-2026-114938), and CERT-FR issued an alert the very next day. It can lead to remote code execution, but only on certain recent x86-64 processors. One notable detail: the researcher who documented it says he spotted it in an hour with a coding assistant, even though a first public report dated back to June.
What Java4Shell fixes
The defect lies in an optimized routine of OpenJDK’s HotSpot virtual machine, the one that speeds up Base64 decoding using the AVX-512 VBMI instructions of x86-64 processors. According to the analysis published by the company Calif, a Base64 string crafted in a certain way skews a length calculation: a 64-byte write starts at position 48 of an array that only holds 50, so 62 bytes overflow past its end. The researchers explain that they turned this overflow, initially hard to control, into a heap memory information leak and then into remote code execution.
In its alert CERTFR-2026-ALE-016, the national agency identifies two risks: denial of service and remote arbitrary code execution. Calif gives the flaw a score of 9.2 out of 10 on the CVSS 4.0 scale, i.e. the “critical” level. The official bulletin does not say whether attacks have already been observed.
A discovery sped up by AI, a report that went nowhere
The timeline published by Calif sheds light on the role coding assistants now play in vulnerability research. Its researcher Quang Luong says he found the defect on 24 October with the help of Codex, OpenAI’s programming tool: about an hour for the discovery, two more hours for a first working exploit, and less than $200 in token usage.
Yet the same document shows that the bug was not unknown. A user who had run into the crash had opened a support ticket with Eclipse Adoptium, which distributes Temurin, as early as 9 June 2026. On 8 September, this user published a proof of concept reproducing the crash; the next day, another contributor proposed a fix to OpenJDK, which was later closed without being merged. Calif passed it on to Oracle on 24 October; the vendor announced a fix was ready on 13 November and released it on the 17th.
What can be drawn from this single case: the assistant did not uncover a completely unknown defect, but it cut the path from that defect to a first exploit down to a few hours. Security teams would therefore do well to deal faster with bugs that have already been publicly reported, even when they look like a simple crash.
Who is actually exposed
Not all Java installations are affected. Three conditions must be met:
- an x86-64 processor supporting AVX-512 VBMI: Calif cites Intel Xeon chips from the Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids generations, as well as AMD EPYC and Ryzen chips from Zen 4 onwards; ARM64 servers are not affected;
- a vulnerable Java version: according to CERT-FR, Oracle lists in particular Java SE 17.0.21.0.3, 21.0.13.0.3, 25.0.5.0.3 and 27.0.1.0.3, a Java 8 variant and several GraalVM editions; according to Calif, OpenJDK is affected from version 17.0.6 onwards, as are the distributions derived from it (Corretto, Temurin, Microsoft, Zulu, Liberica, SapMachine, Red Hat);
- an application that passes attacker-controlled Base64 data to the decoder, a condition highlighted by CERT-FR.
CERT-FR names software widely used in businesses that embeds OpenJDK: Keycloak, Jira, Confluence, JFrog Artifactory, Jenkins, PingFederate, Kafka, Solr, Elasticsearch, Gerrit and Spinnaker. Calif claims to have demonstrated unauthenticated code execution on the first six in their default configuration, and authenticated code execution on the others. As Économie Matin points out, this list serves as a starting point for an inventory: the mere presence of Java on a machine does not prove that it can be attacked.
Steps to take this week
The priority remains applying vendors’ patches. For Temurin, Économie Matin lists versions 17.0.21.1+2, 21.0.13.1+2, 25.0.5.1+2 and 27.0.1.1+4, released on 18 November, and states that Adoptium considers its 8 and 11 branches unaffected. For other distributions, refer to each provider’s bulletin.
In the meantime, two workarounds exist. The startup option -XX:-UseBASE64Intrinsics disables the accelerated routine; the option -XX:UseAVX=2 turns off the use of AVX-512. Calif warns that this second option prevents Java from starting on an ARM machine. These settings change service configurations and must be tested: they do not replace the update.
On Linux, CERT-FR provides a command to read the processor flags: an affected machine shows avx512f, avx512bw, avx512vbmi and bmi2 all at once. Économie Matin also advises checking the Java version actually running on servers and in containers, then testing compatibility before deploying the patch. Finally, services that receive data from the internet, such as an authentication portal, come first.
Frequently asked questions
Is my Windows workstation affected by Java4Shell?
The software cited by CERT-FR consists mainly of tools installed on servers. A workstation is only affected if it has a processor supporting AVX-512 VBMI, a vulnerable Java version and an application that decodes Base64 supplied by a third party. Updating Java remains the simplest measure.
Is the flaw already being exploited by hackers?
CERT-FR’s 18 November alert does not say whether the flaw is already being exploited, and Calif reports no observed attack. However, a proof of concept triggering the crash had been public since September, and Calif says it holds working exploits that it is keeping confidential while updates are rolled out: response time must therefore be short.
Should AVX-512 be disabled on all servers?
No. It is a temporary workaround, to be reserved for machines that cannot be patched right away. It does not work on ARM, where the AVX option prevents Java from starting. The lasting solution remains installing the fixed version provided by the vendor of your Java distribution.
Sources
- CERTFR-2026-ALE-016 — Vulnérabilité dans Oracle Java SE — CERT-FR (ANSSI), 2026-11-18
- Java4Shell (CVE-2026-114938) — Calif, 2026-11-17
- Java4Shell : l'ANSSI alerte, quels correctifs Java pour les entreprises ? — Économie Matin, 2026-11-19