Java4Shell: The AVX-512 Base64 Bug That Lets Strangers Run Code on Your Jira, Jenkins and Keycloak

Java4Shell: The AVX-512 Base64 Bug That Lets Strangers Run Code on Your Jira, Jenkins and Keycloak

Java4Shell: The AVX-512 Base64 Bug That Lets Strangers Run Code on Your Jira, Jenkins and Keycloak

The week's biggest enterprise story is a flaw in the Java runtime itself. Java4Shell, tracked as CVE-2026-114938, is an out-of-bounds write in the Base64 decoder of the x86-64 HotSpot JVM, disclosed by research firm Calif. The bug lives in Java 17.0.6 and later, and the vulnerable code path only runs on Intel and AMD x86-64 chips that support AVX-512 VBMI — but on those machines, specially crafted Base64 input can crash the JVM and allow unauthenticated remote code execution. Oracle and the OpenJDK project shipped fixes on November 17, CISA publicly urged immediate patching on November 18, and as of this writing there is no evidence of in-the-wild exploitation — while Calif rates the flaw 9.2 under CVSS 4.0, Oracle's CVSS 3.1 score is 8.1, and Canada's Cyber Center says it is not aware of any attacks yet.

The trigger is what makes it pre-auth. Servers decode Base64 data — HTTP Basic login headers, SAML artifacts, CSRF tokens — before checking whether a user is logged in, so an attacker with no account can reach the decoder on the very first request. Java uses a faster AVX-512 version of the decoder after its JIT compiler optimizes frequently executed code, and a malformed string ending with an extra = makes that decoder write 62 or 63 bytes past the end of its output buffer. Java rejects the string — but only after the write. The overflowing bytes are fixed in value, so the attacker controls only where they land, which is why early readings judged it a crash rather than an exploit. Calif's chain gets further: arrange the server's memory so the overflow expands a neighboring array, use it to write chosen bytes, and use a second array to read memory back — a step the researchers call JavaBleed — and from there reach the parts of Java that run code. Blackpoint Cyber reproduced the JavaBleed read in 30 out of 30 local runs and crashed a test Spring Boot app with roughly 28,300 malformed login requests over 20 seconds, without logging in; Upwind reproduced the memory corruption on AWS Ice Lake hardware; other labs have so far reproduced crashes but not code execution. Each target needs its own memory layout and timing — there is no universal exploit.

The exposure list is what turns a runtime bug into a fleet problem. Calif's write-up says the pre-auth path works under default settings on Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate, and with a login on Apache Kafka Broker, Apache Solr, Elasticsearch, Gerrit and Spinnaker — with the lists explicitly incomplete. Hospitals were reported racing to patch after CISA's alert, and advisories have poured out from more than two dozen vendors. Red Hat is still preparing fixes for its builds; Amazon has already fixed Corretto. One mitigation note with a familiar shape: a proxy that waits for complete requests before forwarding — as nginx does by default — blocks some but not all of the attacks, and a failed attempt typically just crashes Java, letting the attacker try again.

What Should You Do?

  1. Apply the November 17 Oracle/OpenJDK CPU update everywhere Java 17.0.6+ runs on AVX-512-capable x86-64 hosts — including your JDK vendor builds (Corretto is fixed; check Red Hat, Temurin, Zulu and container base-image status). This is a runtime patch, not an application one: no app upgrade replaces it.
  2. Inventory the pre-auth decoder surface: Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate under default settings are reachable without a login. Anything accepting Base64 request material (Basic auth, SAML, tokens) in front is a candidate.
  3. Use a full-buffering proxy in front of Java services as a partial stopgap — nginx's default request-buffering behavior blocks some attack variants — but treat it as a tripwire, not a fix; CVE-2026-114938 has variants a buffering proxy does not stop.
  4. Monitor for JVM crashes, not just successful RCE. A failed exploit attempt kills the process — clusters of unexplained Java crashes on AVX-512 hardware are the early-warning signal that someone is probing you.

The WAF Angle

Java4Shell is the purest recent example of why WAF coverage has to extend below the application layer: the malicious input is a Base64 string — no SQL metacharacters, no path traversal, nothing a signature engine recognizes as dangerous — and it lands in the JVM's decoder before your application's authentication logic ever runs. Two practical moves come out of this. First, request buffering: a proxy or WAF that fully buffers requests before the origin sees them (nginx's default) defeats the request-splitting variants that depend on the backend consuming partial input — the same full-buffering posture that limits the Apache mod_http2 RCE. Second, rate-based detection: the demonstrated crash path needs tens of thousands of malformed requests in seconds, so a WAF rule that flags request bursts of malformed Base64 on Java endpoints turns an RCE attempt into an alert. Neither replaces the runtime patch — Java4Shell is the PHP2Shell lesson transplanted to the JVM: when the parser is the bug, filter around it, and patch it fast.

Sources