Java4Shell CVE-2026-114938: OpenJDK Base64 RCE on AVX-512 CPUs, what to patch
A heap overflow in HotSpot's AVX-512 Base64 decoder affects OpenJDK 17.0.6 and later on x86-64; here is who is exposed, how to check and the flag that turns off the bad code.
· 8 min read
Written by AI from the sources linked below. How we use AI
On this page
Oracle and the OpenJDK project have a fix for Java4Shell (CVE-2026-114938), a critical memory-corruption bug in the AVX-512 Base64 decoder that HotSpot uses on x86-64 servers. Security firm Calif published its write-up on 17 November 2026 and says a remote attacker with no credentials can use the bug to read heap memory and, in the products it tested, run code. Any Java service that Base64-decodes data an attacker can send, on a CPU with AVX-512 VBMI, is in scope, so the first job today is finding out which of your hosts match.
Key points
- CVE-2026-114938 ("Java4Shell") is a heap buffer overflow in the x86-64 AVX-512 Base64 decoder of OpenJDK HotSpot; the decoder writes 62 bytes past the end of its output array before Java rejects the malformed input.
- It affects OpenJDK 17.0.6 and later (including 21, 25 and 27) on x86-64 CPUs with AVX-512 VBMI; ARM64 hosts such as AWS Graviton are not affected.
- Calif scores it 9.2 (Critical) under CVSS 4.0; Oracle's Critical Security Patch Update of 17 November rates the Java SE entry 8.1 (High) under CVSS 3.1.
- The upstream fix is JDK-8393212, in OpenJDK 27.0.1.1; the 17, 21 and 25 update repositories did not yet carry it on 17 November, so most teams must take their vendor's build when it ships.
- No exploitation in the wild has been reported, and Calif is withholding its weaponised chains for now.
What to do
Who is affected. Every HotSpot release from 17.0.6 onward carries the vulnerable decoder, so this is not only a "Java 17" problem: 21, 25 and 27 are affected too. Endor Labs reports that 17.0.0 to 17.0.5, Java 8 and Java 11 do not have the x86 decode stub. One wrinkle: Oracle's own list, as relayed by Calif, names Oracle Java SE 8u511b33-perf alongside 17.0.21.0.3, 21.0.13.0.3, 25.0.5.0.3 and 27.0.1.0.3 as vulnerable, so check your Oracle Java 8 build rather than assuming. Calif lists Amazon Corretto, Eclipse Temurin, Microsoft Build of OpenJDK, Azul Zulu, BellSoft Liberica, SAP SapMachine, Red Hat OpenJDK, Oracle OpenJDK, Oracle Java SE and Oracle GraalVM for JDK as built from the affected code.
The host must also have the hardware. The vulnerable code path needs x86-64 with AVX-512 VBMI: Intel Ice Lake and later server parts (Sapphire Rapids, Emerald Rapids, Granite Rapids) and AMD Zen 4 or newer. Older Skylake-SP and Cascade Lake Xeons and pre-Zen 4 AMD parts are not affected, per Endor Labs.
How to check. On Linux, Calif suggests looking for the CPU flags avx512f, avx512bw, avx512vbmi and bmi2 in /proc/cpuinfo; all four must be present. Endor Labs adds a check of the running JVM, using the same JDK and options as the service:
grep -o -w -m1 'avx512vbmi' /proc/cpuinfo
java -XX:+PrintFlagsFinal -version | grep -E ' (UseAVX|UseBASE64Intrinsics) '
If the first command prints avx512vbmi, UseAVX is above 2 and UseBASE64Intrinsics is true, the vulnerable path is live on that JVM. To inspect a running process, Horizon3 suggests jcmd <PID> VM.flags -all.
How to fix. Install your Java vendor's update. Oracle shipped its fix on 17 November. Endor Labs reports the Oracle builds as 17.0.21.1 and 21.0.13.1 "reportedly", and says builds for 25.0.5 and 27.0.1 had not been published at the time of its post. Other vendors ship on their own schedules and version labels, so read each vendor's advisory instead of inferring the fixed build from Oracle's numbering. Remember Java that lives inside container images, appliances and bundled products: those need a rebuild or a vendor update too.
Workaround if you cannot patch yet. Calif recommends starting the JVM with either -XX:-UseBASE64Intrinsics, which turns off the hardware-accelerated Base64 code, or -XX:UseAVX=2, which keeps the JVM off AVX-512. Use the second flag only on x86-64 hosts: OpenJDK on ARM does not recognise it and fails to start. Horizon3 points out that Oracle does not publish either flag as a supported workaround, and neither is a fix. Calif also warns that a reverse proxy is not a general mitigation: nginx buffers request bodies by default, which removes the timing some of its exploit chains rely on, but Elasticsearch runs from a single complete request, so buffering does not stop that chain.
What happened
Calif says it found and weaponised the flaw with AI help. In its technical section, researcher Quang Luong describes pointing Codex at HotSpot's JIT compiler, intrinsics, widely reachable APIs and cryptography; most early results were unreachable or unrealistic, and after "around 20 minutes" of prompting through the noise it produced a finding in the Base64 decoder. Calif says it took Luong an hour to discover the bug and two more hours to build the first working exploit, with a total token cost under $200. Those are Calif's own figures and we have not been able to check them independently.
Endor Labs writes that Calif published after the vulnerability leaked ahead of its planned disclosure. Oracle's advisory credits Quang Luong of Calif and a second researcher, Albert Yang, with the report, according to Horizon3.
How it works
To decode Base64 quickly, HotSpot replaces the Java loop with an intrinsic: hand-written x86-64 assembly using AVX-512 VBMI instructions, which the C2 JIT compiler swaps in once the decoder is "hot". The bug sits in the code that handles the final few characters.
The routine builds a 64-bit store mask from the output length. When the input ends in a padded fragment shorter than one four-character block (=, ==, X= and so on), the computed output length is zero, and subtracting the padding underflows it to -1 or -2. The mask function leaves all bits set for lengths of 64 or more, so the mask becomes all ones and the instruction writes a full 64 bytes. Calif's example: 64 A characters followed by one = makes OpenJDK allocate a 50-byte result array, then write 64 bytes starting at offset 48, which is 62 bytes past the end. Java later throws IllegalArgumentException for the bad input, but the neighbouring heap objects have already been damaged.
The shortest trigger depends on the version. Calif gives 65 characters (64 A plus =) for current releases and 129 (128 A plus =) for OpenJDK 17, because the vector path only runs above a length threshold; Endor Labs reports the same thresholds. Calif publishes a bounded proof of concept that writes only inside a real 4,096-byte array and prints REPRODUCED on an affected JDK, so you can test a host without risking a crash.
The overwrite on its own is a fixed value: Endor Labs reproduced it on Temurin 17.0.10 and 17.0.12 under an AVX-512 emulator and found two inputs that differed only in their data characters left identical out-of-bounds bytes. The attacker controls where the write lands, not what it writes, and the usual result is a crash in garbage collection. Calif's route to code execution has stages: grow one array's length, use a second array to leak heap contents (it calls this "JavaBleed"), repair damaged object headers, groom the heap remotely, then reach the JVM's code cache. Calif is explicit that there is no universal exploit: each target needs its own decoder route, writable buffer, feedback channel and heap layout.
The numbers
The two public severity scores differ because they use different scales and come from different parties. Calif's 9.2 is CVSS 4.0; Endor Labs notes no vector string has been published for it. Oracle's Java SE score, CVSS 3.1 with the vector AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H, is 8.1 (Horizon3). Endor Labs says Oracle's November update carried 11 Java SE fixes.
The chart compares two different scoring versions, so read it as "two estimates of a serious flaw", not as a like-for-like disagreement.
What it means
Operators. Exposure is decided by three things together: the Java version, the CPU and whether the application decodes attacker-controlled Base64. That last condition is broad. Base64 appears in authentication tokens, file uploads and network messages. Horizon3 notes internet exposure is not required: any internal service that decodes untrusted Base64 reaches the same path. Calif also describes a client-side exploit from researcher samczsun: a Java program that opens a TLS connection to a malicious server can be compromised even if it exposes no service.
Product teams. Calif says it demonstrated pre-authentication remote code execution in default configurations of Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate, and post-authentication execution in Apache Kafka brokers, Apache Solr, Elasticsearch, Gerrit and Spinnaker. Calif calls its list non-exhaustive. Endor Labs labels these as claims by Calif, with no public proof-of-concept shell, though Calif's page shows a demonstration video. Treat the named products as the first places to patch.
Defenders in general. The disclosure fits a pattern we covered with PHP2Shell, an earlier Calif finding that attackers used against Check Point gateways: CVE-2026-123039 and the PHP2Shell chain. Calif argues the lesson is to delete code you do not need, and that Java's own optimisations ship to every user whether they benefit or not. That is the researchers' view, not an established consensus.
What we don't know yet
- Whether the OpenJDK Vulnerability Group has published an advisory. Calif says OpenJDK released one on 17 November; Endor Labs says it could not retrieve one, and Horizon3 says none has been published. Oracle's own advisory is the clear source.
- The exact fixed builds for each distribution. Endor Labs says Oracle's 17.0.21.1 and 21.0.13.1 are "reportedly" the fixes and that 25 and 27 builds had not been published; Horizon3 lists OpenJDK ranges (before 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1) but flags them as coming from independent analysis.
- Whether the CVE-to-commit mapping is exact. Endor Labs says the mechanism matches Calif's write-up but it could not confirm that CVE-2026-114938 is the CVE for JDK-8393212.
- The CVSS 4.0 vector behind the 9.2 score.
- Whether Java 8 builds beyond Oracle's 8u511b33-perf are affected; Calif's list and Endor Labs differ in how they treat Java 8.
What to watch next
Watch for your vendor's fixed build (Temurin, Corretto, Zulu, Microsoft, Red Hat and Ubuntu or Debian packages), for the OpenJDK advisory, and for Calif's promised release of its product-specific chains once the patching window passes. Also watch for the flaw being added to CISA's Known Exploited Vulnerabilities catalogue; Endor Labs says it was not listed as of 17 November. Until your patch lands, consider the JVM flag on hosts that match all four conditions, and check Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate first.
Sources
- Calif: Java4Shell, Write once, RCE everywhere (primary, the original research)
- Horizon3.ai: CVE-2026-114938, Java4Shell Base64 Overflow (summarises Oracle's Critical Security Patch Update Advisory, November 2026)
- Endor Labs: Java4Shell (CVE-2026-114938), Pre-Auth Heap Overflow in OpenJDK
- Oracle security alerts index
Written by
The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.
Read next
CVE-2026-129448: Apache Roller XML-RPC RCE exploited in attacks, fixed in 6.1.6
Apache says unauthenticated attackers are already exploiting a critical deserialization flaw in Roller 4.0.0 through 6.1.4; here is how to patch or mitigate.
audiotale.net team8 min read
ForgeRock connected-car flaws: researchers unlock cars via identity-platform bugs, no CVE
Two 7 November write-ups show chains from firewall bypass to account takeover and remote unlock on unnamed automakers' ForgeRock logins.
audiotale.net team6 min read
OpenJDK October 2026 advisory: 47 Java CVEs and 6 JavaFX, Corretto 25.0.5
The 20 October 2026 OpenJDK advisory fixes 47 OpenJDK and 6 OpenJFX CVEs, none above CVSS 7.5; here is who is affected and how to update.
audiotale.net team6 min read