Java4Shell, tracked as CVE-2026-114938, is a memory-corruption flaw in the OpenJDK HotSpot runtime. A malformed Base64 string can let an unauthenticated attacker overwrite heap memory. Oracle and the OpenJDK project shipped fixes on November 17, 2026, the same day security start-up Calif disclosed the bug. CISA followed with an alert that urges every organization running Java to update. An AI agent helped find the bug in about an hour, and that should worry defenders as much as the bug itself.
Java runs banking platforms, airline systems, hospital software, telecom stacks and government services. Calif estimates that 90% of Fortune 500 companies use it. This bug sits in the runtime itself, not in a single library. That turns it into a long, messy patch cycle across thousands of products.
What Java4Shell is and where it lives
The flaw is not in Java code at all. It lives in hand-written x86-64 assembly that speeds up Base64 decoding. HotSpot uses this code only on CPUs with the AVX-512 VBMI instruction set. When the JIT compiler spots a hot call to the decoder, it swaps in this fast routine, called an intrinsic.
The intrinsic mishandles some short, padded inputs. A crafted string makes it compute an invalid output length, so it writes past the end of its heap buffer. Java then rejects the input as invalid Base64. By that point, the routine has already damaged the objects next to the buffer. On its own, the damage usually crashes the JVM. With careful heap grooming, an attacker can steer it.
Key facts for triage
- CVE: CVE-2026-114938 (Java4Shell), in the HotSpot compiler.
- Severity: Oracle and OpenJDK score it 8.1 under CVSS 3.1. Calif rates it 9.2 under CVSS 4.0.
- Attack conditions: network access, no login, high complexity, high impact.
- Affected versions: OpenJDK 17.0.6 and later, across Java 17, 21, 25 and 27.
- Affected builds: Oracle Java SE, Oracle GraalVM for JDK, Amazon Corretto, Eclipse Temurin, Microsoft Build of OpenJDK, Azul Zulu and Red Hat OpenJDK.
- Affected CPUs: Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, plus AMD Zen 4 or newer.
- Not affected: ARM64 hosts.
- Fix: Oracle’s November Critical Patch Update and the matching OpenJDK releases, out on November 17, 2026.
Why the CPU matters
Older Xeons without AVX-512 VBMI never take the vulnerable path. So two identical containers can carry different risk, depending on the node that runs them. In a cloud fleet with mixed instance types, a version number alone will not tell you if you are exposed.
How Java4Shell turns a Base64 string into code execution
Modern Java apps decode Base64 everywhere. JSON Web Tokens, SAML assertions, basic auth headers, API payloads and cookies all use it. Many apps decode these values before the user logs in. That puts the bug within reach of anyone on the internet.
The exploit chain
Calif’s chain works in stages. First, the out-of-bounds write corrupts the length field of a nearby buffer. That turns a fixed overwrite into writes of values the attacker chooses. Next, a second corrupted array gives the attacker a read primitive. With it, they can pull private keys and tokens straight from the heap, much like Heartbleed. Code execution follows from there.
Calif reports pre-authentication code execution in default setups of Keycloak, Jira, Confluence, JFrog Artifactory, Jenkins and PingFederate. The team also showed a working break-in against Keycloak. After login, the chain worked against Apache Kafka, Apache Solr, Elasticsearch, Gerrit and Spinnaker. It also hit sample Apache Struts and Spring Boot apps. The researchers say these lists are not complete.
There is a client-side angle too. A malicious server can compromise a Java program that opens a TLS connection to it. That brings build agents, data pipelines and outbound integrations into scope, not only public web apps.
Limits of the attack
Exploitation is not trivial. Each target needs its own heap profile. Network latency and jitter can break the grooming, and a failed attempt usually crashes the JVM. No universal payload exists. At the time of writing, nobody had reported attacks in the wild, and CISA had not added the CVE to its KEV catalog. Calif is holding back app-specific exploit chains to give teams time to patch. That window will not stay open for long.
Why an AI-found Java flaw changes the patching clock
How Calif found the bug may matter most. A researcher pointed an AI coding agent at the JIT compiler, its intrinsics and the APIs around them. The agent surfaced the flaw in about an hour. A working exploit took roughly two more hours, and the compute bill came in under $200.
This is the second major runtime bug that AI has helped uncover in a few weeks. In October, researchers disclosed decade-old flaws in PHP. Some companies and governments took websites offline as a precaution. The pattern is clear. Teams long assumed that the native code inside managed runtimes was well reviewed. Now anyone can audit it cheaply and at scale.
That cuts both ways. Attackers have the same tools Calif used. They can diff the patch and rebuild the exploit chain on their own. A complex memory bug no longer buys defenders weeks before weaponization. Plan as if a public exploit for at least one common product will appear soon.
The timing adds pressure. The alert lands during a run of major outages across US services. Calif calls Java critical infrastructure, and few would argue. Even without active attacks, boards and regulators will want a clear answer on exposure within days.
What to do now about Java4Shell
- Inventory every JVM, not just every app. Java hides in containers, appliances, CI agents, Kafka and Elasticsearch clusters, and vendor products with their own JRE. Use SBOMs, image scans and host discovery of
javaprocesses. - Map exposure by CPU. On Linux, look for
avx512vbmiin/proc/cpuinfo. Patch x86-64 nodes with that flag first, starting with public and identity services. - Patch to the November 2026 builds. Upgrade Oracle Java SE, GraalVM and your OpenJDK build to releases from November 17, 2026 or later. For products with an embedded JRE, wait for and apply the vendor’s update.
- Mitigate where you cannot patch yet. The flag
-XX:-UseBASE64Intrinsicsturns off the fast Base64 routine. On x86-64 only,-XX:UseAVX=2turns off AVX-512. Keep the AVX flag off ARM hosts, where it can stop the JVM from starting. Test both for performance. - Restart after patching. A running JVM keeps the old code until the process restarts. Check the version in each live process, not just on disk.
- Hunt for warning signs. Watch for JVM crashes and
hs_err_pidfiles. Look for bursts of malformed Base64 in tokens, headers or SAML requests. Flag odd outbound traffic from Java services. - Rotate secrets if you find anything. The read primitive can leak keys and tokens. If you suspect a breach of Keycloak, PingFederate or Jenkins, rotate signing keys, service credentials and session secrets.
Nobody has reported Java4Shell attacks in the wild yet, and exploitation takes real effort. But the bug lives in a runtime most enterprises depend on. Attackers can reach it before login, and an AI agent found it in an afternoon. The teams that come through cleanly will know where every JVM runs. They will patch and restart them this week.


