Java4Shell (CVE-2026-114938): The first 24-Hours: Detections Available Today
Written by detections.ai · Organizer

Within 24 hours after advisories on the vulnerability were published, rules on detections.ai for Java4Shell (CVE-2026-114938) are already starting to surface. As with many exploitation-style CVEs, we’re seeing strong correlating overlap from this vulnerability and previous Java exploitations.
We identified dozens of older detection rules that match what we expect a working exploit to do next, some of them written for Log4Shell in December 202. The vulnerability was disclosed on 17 November, with no reported attacks and no verified public exploit for code execution.
Threat overview: a 62-byte overwrite
Calif's example trigger for Java4Shell is 64 letter A's followed by an equals sign, and on Java 17 it takes 128. Calif, the firm that disclosed the bug, says the JVM writes 62 bytes past the end of the output array while decoding it. Java then rejects the string with an IllegalArgumentException, which Equixly says is the exception a patched JVM throws for the same input.
CVE-2026-114938 is a heap overflow in the Base64 decoder of HotSpot, the JVM in Oracle Java and OpenJDK builds. The flaw is in an AVX-512 routine that the JIT compiler switches to once the decoder has run often enough. OpenJDK lists Java 17, 21, 25 and 27 as affected, and Oracle adds its Java 8 performance build and GraalVM. Both score it 8.1 on CVSS 3.1, with attack complexity high.
OpenJDK's advisory does not list Java 8 or 11 as affected. Only x86-64 processors with AVX-512 VBMI run the routine, according to Calif, Red Hat and Google, and Calif and Amazon Linux say ARM64 is not affected. Calif names Intel Xeon from Ice Lake to Granite Rapids and AMD Zen 4 and newer.
Calif, which says it found the bug with OpenAI's Codex in about an hour, claims remote code execution without authentication against default installs of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate, and is withholding those exploits. Wiz says its AI harness built a proof-of-concept exploit within two hours, and VulnCheck says a full exploit is available to its Initial Access Intelligence customers. Neither is public.
Upwind and Datadog reproduced the overwrite and a JVM crash, and Upwind also built heap read and write primitives. VulnCheck crashed a service with a burst of requests, so denial of service needs no exploit chain. Equixly, writing before Calif's write-up, called code execution unproven.
No source is currently reporting exploitation
CISA issued the alert on 17 November without adding the CVE to its Known Exploited Vulnerabilities catalog, and Canada's Cyber Centre said it was not aware of active exploitation.
Oracle's fixed builds are 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1, and Amazon Corretto, Azul Zulu and SapMachine published fixed builds on 17 November. At the time of publication, Eclipse Temurin, the Red Hat build of OpenJDK along with Microsoft's build and BellSoft Liberica had not. Until a fixed build is running, Red Hat prefers starting the JVM with -XX:-UseBASE64Intrinsics, and SonarSource applies that flag to SonarQube.
Exposure depends on the running JVM, the processor under it, and whether untrusted input reaches Java's Base64 decoder. Check a process with jcmd <pid> VM.version and jcmd <pid> VM.flags -all, and the host for the avx512vbmi flag in /proc/cpuinfo. CISA, Qualys and Endor Labs warn that a JDK bundled inside a product or a container image may be missing from an inventory, and a container keeps its old flags until it is recreated.
Look for a Java service that starts a shell or a download tool, or writes a script file. Treat a JVM crash on an exposed host as a reason to check that host.
Detections available today
Rules on detections.ai specifically targeted to Java4Shell, or the associated CVE are already trickling in.
While nobody has published telemetry from a working exploit, in earlier Java vulnerabilities, code execution has tended to show up as the service starting a shell or a download tool, or writing a script file, and public rules for that go back at least to Log4Shell. We recommend starting there.
Endor Labs says Calif's demonstration video shows web shells planted on the exploited products. We read about 120 of the 848 public rule documents whose query mentions Java and found 26 rules that match that behavior. Splunk dates its version 13 December 2021, and SigmaHQ's Windows rule is dated four days later.
The 26 are titled for the behavior or for earlier bugs in Tomcat, PaperCut, XWiki, Confluence and Ivanti products, so a search for the new name does not lead to them. Fifteen people searched on 17 November for the name, the CVE id or a generic term such as java, and none of them opened one of the 26, which is a small sample. Three other people opened Splunk's rule that day.
We hand-picked 10 notable detections, and these are the ones we would deploy first, in kill-chain order. All ten key on behavior and fire after code execution, so the order follows an attacker's first steps on the host: a shell started, a process the server has not run before, a web shell written. You need a free login to see a rule's query, and the SigmaHQ and Splunk rules are also in those projects' public repositories.
T1059.004: Java4Shell CVE-2026-114938 Repeated JVM Crashes Followed by Shell Spawn
YARA-L · Restarts and a shell · Needs process launch events in Google SecOps
On one host within 30 minutes: three or more launches of java or javaw, which stand in for a JVM that keeps crashing and restarting, and a Java process starting sh, bash, dash, zsh, ksh or cmd.exe. The query does not check which came first.
This is our rule, written on 17 November from the first reporting. It needs the launches and the shell, so it fires no earlier than the rules below and misses an exploit that works without a crash. Jenkins and other CI agents, and wrapper scripts that restart a JVM, will match.
Web or Application Server Spawning a Shell
SPL · Shell started · Needs the Endpoint.Processes data model
A web or application server process starting a shell, on Linux or Windows. A java or javaw parent counts only when its command line carries a server marker, and the markers include Tomcat, Jetty, Spring Boot, Quarkus, Keycloak, Jira, Elasticsearch, Gerrit, Solr and jenkins.war. Splunk dates the rule's creation 13 December 2021 and its last modification 28 October 2026.
The shell lists are in Splunk's linux_shells and windows_shells macros. Splunk says developer build systems may need filtering. The marker list has no entry for PingFederate. Splunk's older Linux Java Spawning Shell matches any java or tomcat parent, with no marker.
Suspicious Child Process Spawned by Java on Linux
Sigma · Shell or download tool started · Needs Linux process creation
A parent image ending in /java and a child command line containing a shell name (/bin/sh, bash, dash, ksh, zsh, csh, fish) or curl, wget or python. SigmaHQ rates it high and dates it 3 June 2022.
It has no exclusions, and it matches a substring anywhere in the command line. Build tools and deployment scripts that shell out from Java will match.
Shell Process Spawned by Java.EXE
Sigma · Shell started · Needs Windows process creation
A parent image ending in \java.exe starting bash.exe, cmd.exe, powershell.exe or pwsh.exe. It skips events where the parent path and the command line both contain build. SigmaHQ dates it 17 December 2021 and cites a Log4Shell write-up.
A javaw.exe parent is not matched, and neither is a service that hosts the JVM in its own executable, such as tomcat9.exe.
Suspicious Web Application Child Process Execution
XQL · Shell or download tool started · Needs Cortex XDR process events from Linux agents
A process named java, tomcat, catalina.sh or jetty, or one of ten other web and application servers, starting a shell, curl, wget, nc, socat, perl, python, chmod, base64 or busybox.
It has no exclusions. The rule lists maintenance scripts and cron jobs run under the web server account as false positives.
Web Shell Execution via Web Server Spawning Command Interpreter
S1QL · Shell or system tool started · Needs SentinelOne process creation events
java.exe, javaw.exe or one of eight other Windows web server processes starting any of 20 interpreters and system tools, among them cmd.exe, powershell.exe, certutil.exe, whoami.exe and rundll32.exe.
It excludes parents published by Microsoft Corporation. Check whether that also drops java.exe from the Microsoft Build of OpenJDK on your hosts.
PaperCut RCE (CVE-2026-82078): PaperCut Server Spawning Shell Process
CQL · Shell started · Needs CrowdStrike ProcessRollup2 events from Windows
A parent named java.exe, pc-app.exe, pc-tomcat.exe or PCClient.exe starting cmd.exe, powershell.exe or pwsh.exe. It was published in September 2026 for a PaperCut bug, and the java.exe parent makes it match any Windows service that runs as java.exe.
It has no exclusions. Agents and management tools that run scripts through a Java service will match.
Shell or downloader spawned by Confluence/GitLab server processes (C3RB3R)
KQL · Shell or download tool started · Needs DeviceProcessEvents
java, java.exe or javaw.exe starting a shell, curl, wget or certutil.exe, where the parent's command line or folder path contains confluence, atlassian, gitlab or tomcat. It looks back 7 days and also covers GitLab's Ruby and Perl processes.
It skips parents whose command line contains jenkins, bamboo or teamcity, so it stays quiet on CI hosts and does not cover Jenkins. A Java service whose path and command line carry none of the four names is not matched.
Unusual Web Server Child Process Creation
KQL · New child process · Needs DeviceProcessEvents and DeviceInfo
On internet-facing Windows servers, a parent and child pair under a web server process that has not appeared on that device in the previous 30 days. The baseline means it needs no list of shells or tools, and the rule's description says to run it daily.
It treats java.exe as a web server only when the command line carries Tomcat's Catalina bootstrap class. A Java service on another server is not covered.
Java Process Writing JSP File
SPL · Web shell written · Needs the Endpoint.Processes and Endpoint.Filesystem data models
A process named java, java.exe or javaw.exe, joined on process GUID to a file write whose name contains .jsp. Splunk wrote it for Spring4Shell.
Application deployments and updates write JSP files too. Splunk says to filter the ones that are normal for your environment.
Hunting the stages before code execution
Currently, each public rule is at the last stage, after the attacker has achieved code execution on the exploited server. The stages before it are recorded in application logs, proxy logs and the JVM's working directory, and a rule that reads process telemetry sees none of them.

27 public rules on detections.ai as of 18 Nov, 01:25 UTC, counted once each and placed by what the query looks at. One of the 27 also requires repeated Java launches. The shell and the web shell are what we expect after code execution, and no source has observed them. Stage labels are ours.
Most of the first public signatures for Java4Shell look at the request. Upwind's check is a Base64 token that contains an = and whose length is not a multiple of four. Equixly published a WAF regex for the Basic authentication header, and a community repository, ikow/java4shell-poc, has Suricata, Sigma and ModSecurity rules for the same check, with no statement that they were tested.
Those rules need decrypted HTTP traffic, and the Basic authentication header is one of several paths to the decoder. HeroDevs traced it to Basic authentication in Spring Security, Jetty and recent Tomcat versions and to Auth0's java-jwt library, and Wiz says WAF rules are unlikely to cover every path. We would run request signatures as hunts.
Two public tools announce themselves in the user agent: a version scanner as java4shell-scanner/1.0 and a trigger harness as java4shell-remote/1.0. Both strings are worth a hunt in proxy logs until someone changes them.
These are the places an attempt can leave a trace before a shell appears.
A run of Base64 decode failures from one source
The request · Application logs · No public rule
Calif's exploit first sends many valid Base64 values so that the JIT compiles the fast path, and then the malformed input. Where the application logs it, the failure reads IllegalArgumentException: Input byte array has wrong 4-byte ending unit, and Upwind saw that line on every attempt in its lab. A busy service has already compiled the fast path, so the valid run may be missing. One such line is ordinary bad input, but a run of them from one source against an authentication endpoint is worth a hunt.
Equixly says a patched JVM throws the same exception, so this finds attempts and cannot tell you whether one worked.
A JVM crash on an exposed host
Crash · JVM working directory, syslog · No public rule
Calif says a miss can crash the JVM and that in many instances it can try again. Upwind's crash was a SIGSEGV in a G1 garbage-collection routine, which leaves an hs_err_pid<n>.log file in the working directory. None of the 55,224 public rule documents on detections.ai mentions hs_err_pid. Outside the platform, ThreatDefence has example hunts for the crash file and the JVM's fatal-error line in syslog, and the community repository has experimental Sigma rules for the crash file and repeated decoder exceptions.
Crash rates vary by test. VulnCheck says one request usually does not crash a default server and a burst does, and Equixly, reasoning from the source code, expects most attempts to change nothing.
Finding the other rules on detections.ai
Start from the Threat Report, Java4Shell Critical Heap Overflow Vulnerability.
AI Assistant can draft a detection from the Threat Report in your query language, and the draft needs testing like any rule above. On the enterprise tier, Nightwatch hunts your connected telemetry for the activity in each new report and checks your own detection library for coverage.
For the stages with no public rule, ThreatDefence's five example hunts, seven KQL queries in all, cover inventory, inbound connections to Java, the crash file, the syslog line and new destinations. It published them before the disclosure, and we have not tested them.
Then search for the behavior, for example java spawning shell or web shell, because none of the older rules carries the new name. Log4Shell is the closest precedent for what happens after exploitation.
Appendix
Additional public Java4Shell detections
• YARA-L T1059.004: Java4Shell CVE-2026-114938 Repeated JVM Crashes Followed by Shell Spawn (fires on: JVM restarts and a shell; parent: any Java process; written for Java4Shell)
• SPL Web or Application Server Spawning a Shell (fires on: shell start; parent: Java with a server marker; written for Log4Shell)
• Sigma Suspicious Child Process Spawned by Java on Linux (fires on: shell start; parent: any Java process)
• Sigma Suspicious Processes Spawned by Java.EXE (fires on: download and system tools start; parent: any Java process; written for Log4Shell)
• Sigma Java/Tomcat Process Spawning Discovery Commands (Fileless Web Shell) (fires on: shell start; parent: any Java process)
• Sigma Java process spawns shell/cmd child (log4j FOIS deserialization RCE) (fires on: shell start; parent: java.exe or javaw.exe; skips Maven, Gradle, Jenkins and catalina.sh parents; written for log4j deserialization)
• Sigma JDWP Exploitation - Suspicious Script Download via Java Process (fires on: download tool; parent: any Java process; written for JDWP)
• KQL XWiki CVE-2025-24893 Java process spawning bash /dev/tcp reverse shell (fires on: reverse shell; parent: any Java process; written for XWiki CVE-2025-24893)
• KQL Apache Tomcat — Java deserialization reverse shell pattern (CVE-2026-34486) (fires on: reverse shell; parent: any Java process; written for Tomcat CVE-2026-34486)
• KQL Shell or downloader spawned by Confluence/GitLab server processes (C3RB3R) (fires on: shell or download tool start; parent: Java with confluence, atlassian, gitlab or tomcat in its path or command line; written for Confluence, GitLab)
• YARA-L Ivanti Web Server Shell Spawning (fires on: shell start; parent: any Java process; written for Ivanti)
• YARA-L Most Important ShinyHunters Detection 2026: PeopleSoft App Server Spawning Shell After Inbound HTTP (CVE-2026-35273) (fires on: shell start within 5 minutes of an inbound connection; parent: any Java process; written for PeopleSoft CVE-2026-35273)
• YARA-L Potential Webshell Process Execution from Web Servers (fires on: shell start; parent: Tomcat only)
• S1QL Web Shell Execution via Web Server Spawning Command Interpreter (fires on: shell start; parent: any Java process)
• XQL Suspicious Web Application Child Process Execution (fires on: shell start; parent: any Java process)
• CQL PaperCut RCE (CVE-2026-82078): PaperCut Server Spawning Shell Process (fires on: shell start; parent: any Java process; written for PaperCut CVE-2026-82078)
• SPL Java Process Writing JSP File (fires on: script write; parent: any Java process; written for Spring4Shell)
• Sigma Suspicious File Write to Webapps Root Directory (fires on: script write; parent: Tomcat only)
• KQL Web shell: .aspx/.jsp file created in web app path by non-admin process (fires on: script write; parent: Java not running as a system account)
Limits and sources
• That a working exploit starts a process is our expectation, and ThreatDefence and Qualys share it. CISA says only to watch for suspicious Java processes, and no source describes what an attacker did after a real Java4Shell exploit.
• Calif and VulnCheck claim code execution and Wiz claims a proof-of-concept exploit. Oracle rates attack complexity high and says no privileges are needed. Amazon Linux scores the bug 8.8 with low privileges required and describes code execution by an authenticated user, and NCSC-NL says exploitation is not easy because of several prerequisites. Calif publishes a CVSS 4.0 score of 9.2 without a vector, and it is not in the CVE record.
• Fix status and vendor advisories were last checked at 01:13 UTC on 18 November, and will have changed.
• The search figures cover only 15 people. They come from logged-in members with detections.ai staff excluded, and nobody searched for the name or the CVE id on the logged-out discovery page.
Rules: published rules in community groups, counted once per distinct rule. A rule document is one copy of a rule in one group. A rule names Java4Shell when the word or the CVE id is in its title, description, logic or CVE field. The 26 older rules were found by reading public rules that mention java, java.exe or javaw and keeping those whose query matches a Java process starting a shell or a download tool, or writing a script file. Search: distinct people whose query was java4shell or contained 114938; the generic terms were java, jvm, jdk, openjdk and java remote code execution. Usage: distinct people in product analytics, matched to rules by id. No figure identifies a member.
Research: Calif, Oracle, OpenJDK, CISA, Upwind (blog, microsite), Equixly, watchTowr, Wiz, VulnCheck, Endor Labs, Datadog Security Labs, Qualys, ThreatDefence, HeroDevs, Red Hat, Google Cloud, Amazon Linux, Azul, SonarSource, Canada's Cyber Centre, NCSC-NL.
Last updated 18 November 2026, 01:35 UTC. Platform figures are as of 18 November 2026, 01:25 UTC. This page will be updated as new rules are published.


