Java’s Base64 decoder has a fast path, and that fast path is where the problem lives. On x86-64 processors that support AVX-512 VBMI, a malformed Base64 string ending in an extra = makes the optimized decoder write 62 or 63 bytes past the end of its output buffer. Java rejects the string — but only after the write has already happened.
On its own, that sounds like a crash. Calif, the security research firm that disclosed the issue this week, says it is enough to reach remote code execution on Jira, Jenkins, Keycloak and three other widely used applications without any login at all. The flaw is tracked as CVE-2026-114938, and Oracle and the OpenJDK project shipped fixes today.
What Java4Shell actually is
The bug sits in the Java compiler component and affects the Java 17 line from 17.0.6 onward, plus Java 21, 25 and 27. Classic Java 8 and Java 11 builds do not contain the vulnerable code, with one exception: the Java SE 8 Performance Pack build 8u511b33-perf. Oracle’s CVE record also lists GraalVM for JDK 17.0.21.0.3 and 21.0.13.0.3, and GraalVM Enterprise Edition 21.3.20.0.3, as affected.
The second condition is hardware. The fast decoder only runs on x86-64 chips with AVX-512 VBMI support: Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD Zen 4 or newer. ARM64 hosts do not execute it. That detail matters operationally, because the same container image can be vulnerable on one server and safe on another.
Why the decode runs before the login check
Servers routinely decode Base64 before they know who is asking. HTTP Basic authentication headers, SAML artifacts and CSRF tokens all arrive encoded, and the application has to decode them to decide whether the request is legitimate. Java compiles the fast version of the decoder for code paths that execute often, so on the affected hardware the vulnerable routine is the default one in exactly the place where unauthenticated input arrives.
Calif says the attacks worked without authentication under default settings on Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate. On Apache Kafka Broker, Apache Solr, Elasticsearch, Gerrit and Spinnaker, a login was required. The researchers describe both lists as incomplete.
Turning fixed bytes into controlled memory
The overflow writes fixed bytes, so an attacker does not directly control what lands past the buffer — only where. That is why early source-code readings pointed at a denial of service rather than code execution.
The researchers’ answer is a heap layout trick they call JavaBleed. By arranging the server’s memory so the overflow expands a neighbouring array, they gain an array whose length they can influence, and use it to write bytes of their choosing. A second array lets them read memory back. From there the attack reaches the parts of the JVM that execute code. There is no universal exploit: each target needs its own memory layout and timing, and some attempts are sensitive to network latency.
What has been proven, and what has not
Independent labs have reproduced the corruption but not the full chain. Upwind triggered the memory corruption on an AWS instance with an Intel Ice Lake chip, damaging the next object in memory and eventually crashing Java. Blackpoint Cyber’s Adversary Pursuit Group crashed a test Spring Boot application by sending roughly 28,300 malformed login requests over about 20 seconds, without ever authenticating. The group also ran the JavaBleed step locally and read a secret stored next to the overflow in all 30 attempts, but did not reproduce remote code execution.
Scores differ sharply depending on who is doing the scoring. Calif rates the flaw 9.2 under CVSS 4.0. Red Hat’s preliminary score is 9.8 under CVSS 3.1. Oracle rates it 8.1 and calls it difficult to exploit, while still noting that a successful attack can result in takeover of Oracle Java SE. Amazon’s entry describes a JVM crash for unauthenticated users, or code execution by an authenticated low-privileged user in certain application environments. Canada’s Cyber Centre says it was not aware of any exploitation, and the flaw was not in CISA’s KEV catalogue when The Hacker News checked.
What to do next
- Patch the JDK this week. Oracle and OpenJDK have fixes out; Amazon has patched Corretto; Red Hat is still preparing its builds. Treat vendor timing as the constraint, not the disclosure date.
- Check whether your hosts can run the vulnerable path at all. On Linux, run:
grep -m1 '^flags' /proc/cpuinfo | tr ' ' 'n' | grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'— all four flags must appear. - Audit the applications, not just the runtime. Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate are the unauthenticated cases Calif named.
- Keep request buffering in front of the app. nginx waits for complete requests by default, which blocks some of these attacks — but not all of them.
- Watch for crash loops. A failed attempt can kill the JVM, and the attacker can simply try again, so repeated unexplained restarts are a signal worth escalating.
Sources
The Hacker News, “Java4Shell Flaw (CVE-2026-114938) Enables Pre-Auth RCE in Jira, Jenkins, Keycloak and More”; NVD entry for CVE-2026-114938 (Oracle); Qualys, patching guidance for the same CVE.



0 Comments