Java4Shell CVE-2026-114938: pre-auth RCE in Java's Base64 decoder

Oracle shipped its November 2026 Critical Security Patch Update on November 17, and one bug inside deserves its own emergency: CVE-2026-114938, nicknamed Java4Shell, an out-of-bounds heap write in the HotSpot Base64 decoder. Security firm Calif released the full write-up and PoC ahead of the coordinated disclosure, and CISA put out a same-day alert urging immediate patching. No exploitation in the wild has been confirmed and the bug is not on CISA’s KEV catalog as of the November 17 feed. CISA didn’t wait for the catalog, and neither should you: I’d patch every Java 17+ runtime this week.
The trigger is a 65-character Base64 string: 64 A’s and an equals sign (128 A’s on older OpenJDK 17 builds). HotSpot compiles the Base64 decode path into an AVX-512 intrinsic, and the bug sits in the hand-rolled assembly that handles the padded tail: the store mask is miscalculated, the routine computes a negative output length, and a full 64-byte store lands 62 bytes past the end of a 50-byte result array. Java’s own validation rejects the input, but the check runs too late. The corruption alone is a crash. Calif’s write-up turns it into a Heartbleed-style memory read and then code execution through remote heap grooming. samczsun’s variant is nastier: a Java client that merely opens a TLS connection to a malicious server gets compromised, service or not.
Scope: OpenJDK 17.0.6+ across the 17, 21, 25 and 27 lines, GraalVM for JDK 17.0.21.0.3 and 21.0.13.0.3, GraalVM Enterprise 21.3.20.0.3, and the Java 8 Enterprise Performance Pack build 8u511b33-perf. Standard Java 8 and 11 skip the intrinsic; ARM64 hosts and x86 hosts without AVX-512 VBMI are not known affected. Intel Ice Lake through Granite Rapids and AMD Zen 4 or newer are the exposure, and a /proc/cpuinfo check for all four of avx512f, avx512bw, avx512vbmi and bmi2 settles the CPU question. Corretto, Temurin, Microsoft, Azul Zulu, Liberica, SapMachine and Red Hat all build this code, and since the bug lives in the JDK, dependency scanners will fly right past it. Calif verified pre-auth RCE in default configs on Keycloak, Jira, Confluence, JFrog Artifactory, Jenkins and PingFederate, and post-auth RCE on Kafka, Solr, Elasticsearch, Gerrit and Spinnaker.
Fixes shipped with Tuesday’s CSPU and the same-day distributor advisories; Red Hat carries it as RHSB-2026-018 and OpenJDK published its own. Pull your vendor’s current build and restart every JVM, including the ones behind the VPN. If you can’t patch before morning, two JVM flags stopped the overflow in VulnCheck’s testing on OpenJDK 17, 21 and 25: -XX:-UseBASE64Intrinsics (pair it with -XX:+UnlockDiagnosticVMOptions) or -XX:UseAVX=2. The AVX switch is x86-64 only, since Java on ARM refuses the flag at startup. Bridges, not the plan.
On scores, Oracle’s record says 8.1 with attack complexity High, Red Hat says 9.8, and Calif’s CVSS 4.0 says 9.2. The disagreement is about how precise the heap work has to be; six demonstrated pre-auth chains say it is routine.
My read: an OpenJDK fix proposal carrying this exact trigger went up on September 9, sat in the queue, and was closed by an inactivity notice in early November, days before the leak (The Hacker News has the timeline). An optimization almost nobody needed ships to every JVM, and the tradeoff stays theoretical until a PoC lands in a queue nobody watches. Patch, restart, and know which of your JVMs decode untrusted Base64.
Questions or war stories: info@stackref.com.
This post was created with a Hermes AI agent (hermesagent.ai) from a weekly news gathering workflow, then reviewed and approved by Keith before publication.

