A malformed Base64 string makes OpenJDK write up to 62 bytes past the end of a heap array, on x86-64 CPUs with AVX-512 VBMI. The floor is a reliable remote crash with no credentials; the ceiling is arbitrary heap read and write.
What happened
Oracle shipped the fix in its November 2026 Critical Patch Update on 17 November, alongside eleven Java SE patches. The OpenJDK Vulnerability Group published a matching advisory the same day. Upstream issue is JDK-8393212.
The mechanism. HotSpot replaces hot Java methods with hand-written machine code called an intrinsic. Intrinsics are fast and they sit outside Java's bounds checking, so each one has to get its own bounds right. Base64.getDecoder().decode() is one of them, and on CPUs with AVX-512 VBMI the JVM uses a vectorised stub that processes 64 characters per step.
decode() runs in three steps: size the output array, decode into it, then validate the input. Feed it an invalid length ending in =, such as 64 Q characters followed by =, and the stub's padding path computes a negative output count. That negative turns into a mask enabling all 64 bytes, and the stub performs a full 64-byte store at the end of an array that is too short. Up to 62 bytes land past the end. Java throws IllegalArgumentException afterwards, which is to say after the write.
Why a fixed overflow still escalates. The overflow bytes come from the stub's lookup table and are always the same pattern, so the attacker does not choose them. The attacker does choose the output array size, which decides where the allocator places it and therefore what sits next to it in memory. Java packs objects with no gap, so the overflow lands on the next object's header.
If that is an ordinary object, the header is corrupted and the JVM crashes the next time anything touches it. That is a repeatable remote crash on every node, reachable with no account. If it is an array the attacker placed there, the length field becomes roughly 1.7 billion and indexing it reads and writes across the heap.
Severity. Oracle is the assigning CNA and scores it 8.1 High under CVSS v3.1. Calif scores it 9.2 Critical under CVSS v4.0
Impacted packages and versions
The flaw is in the runtime. No application or library has to be at fault.
Release line | Affected | First fixed (OpenJDK) |
Java 27 | before 27.0.1.1 | jdk-27.0.1.1-ga |
Java 25 | before 25.0.5.1 | jdk-25.0.5.1-ga |
Java 21 | before 21.0.13.1 | jdk-21.0.13.1-ga |
Java 17 | 17.0.6 to before 17.0.21.1 | jdk-17.0.21.1-ga |
Java 17.0.5 and earlier | not affected | n/a |
Java 11 | not affected | n/a |
Java 8 | not affected, except Oracle Java SE 8 Enterprise Performance Pack 8u511b33-perf | 8u513-perf |
Amazon Corretto fixed builds: 17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2, 27.0.1.1.4.
Am I affected?
All three conditions have to hold at once. A scanner matching package versions sees the first and cannot see the other two.
1. The runtime is an affected build. Check what the application executes, not the system package. A JDK inside an image does not receive host updates.
2. The CPU exposes AVX-512 VBMI. Intel Ice Lake, Sapphire Rapids and later; AMD Zen 4 and later. Most current cloud instance families have it. ARM is not affected. Skylake-SP and Cascade Lake have AVX-512 without VBMI and are not affected.
3. Attacker-controlled Base64 reaches the decoder before validation. The affected path is Base64.getDecoder() and Base64.getUrlDecoder(). The MIME decoder did not reproduce. Common pre-auth paths are Authorization headers, session and bearer tokens, cookies, JSON fields and certificates.
Most teams cannot answer this from their own code, because the decoding happens in software they did not write. Calif reports pre-auth reachability in six widely deployed applications including Jira, Jenkins and Keycloak.
Resolution
CleanStart images
Fixed and available:
Image | Fixed version |
|---|---|
| 17.0.21.1.2 |
| 27.0.1.1+4 |
CVE IDs, CWE and references
Field | Value |
CVE | CVE-2026-114938 ("Java4Shell") |
CWE | CWE-122, heap-based buffer overflow |
Upstream issue | JDK-8393212 |
Oracle component | Compiler, in Oracle Java SE and GraalVM for JDK |
CNA | Oracle |
Reserved / published | 22 October 2026 / 17 November 2026 |
CVSS v3.1 | 8.1 High, Oracle. CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H |
CVSS v4.0 | 9.2 Critical, Calif. Vector not published |
NVD / EPSS | No independent score; no EPSS score, as of 18 November 2026 |
CISA KEV | Not listed as of 18 November 2026 |
