Java4Shell Lets Malformed Base64 Overwrite the Heap in OpenJDK on AVX-512 Servers, but Vendors Rate It From 8.1 to 9.8, the CVE Text Names Neither Base64 Nor the CPU, and Eclipse Temurin and Ubuntu Had No Fixed Build Hours After Oracle’s Patch
Calif’s Java4Shell flaw is a four-line fix in OpenJDK’s AVX-512 Base64 decoder, but exposure, severity and fixed builds differ by source, so here is what each record says and what to check.
Calif published its write-up of “Java4Shell” on November 17, the day Oracle shipped a fix in its November Critical Security Patch Update and CISA told organizations to patch immediately. The bug, CVE-2026-114938, sits in the hand-written x86-64 AVX-512 code that OpenJDK uses to decode Base64. Calif says it demonstrated pre-authentication code execution in default configurations of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate, and SecurityWeek’s headline read “Java4Shell Vulnerability Puts Many Enterprise Applications at Risk of Takeover”.
Table Of Content
- At a glance
- What the bug is, in the code
- Who is actually exposed
- The CVE entry lists none of this
- One bug, four scores
- Is anyone exploiting it?
- Which builds are fixed
- The timeline in the public records
- What to check on your own servers
- Find every runtime, including bundled ones
- Check whether the hardware can take the vulnerable path
- Do not read UseBASE64Intrinsics as an exposure test
- If you cannot patch yet, pick the flag that fits your fleet
- Prioritize authentication paths, and do not rely on size limits
- What we could not confirm
- Sources
The primary records draw a narrower picture, and they disagree with each other in useful ways. The flaw needs one family of CPUs, every exploit chain is built per application, vendors rate it anywhere from 8.1 to 9.8, the CVE text names neither Base64 nor the hardware, and when we checked between about 10:10 and 10:20 UTC on November 18, roughly 13 hours after Oracle’s patch, Eclipse Temurin and Ubuntu had no fixed build. This report sets out what each source says, what the code shows, what we tested on one machine, and what to check on your own servers. Calif is also the firm behind PHP2Shell, which we covered in one report and its follow-up.
At a glance
- What: an out-of-bounds write in the AVX-512 code path of
java.util.Base64decoding, fixed by a four-line change under JDK-8393212. - Who is exposed: HotSpot JVMs from 17.0.6 on x86-64 CPUs with AVX-512 VBMI (Intel Xeon from Ice Lake on, AMD Zen 4 and newer). ARM64 is not affected, and the OpenJDK matrix marks no Java 8 or 11 release.
- Fixed upstream in: 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. Oracle, Azul and Amazon have shipped fixes. At our check Temurin and Ubuntu had none, and neither Microsoft’s release notes nor Red Hat’s bulletin listed one.
- Exploitation: none reported by any source we read. A bounded trigger is public; the per-product chains are withheld.
- Interim mitigation:
-XX:-UseBASE64Intrinsics, which exists on every platform.
What the bug is, in the code
Calif places the flaw in HotSpot’s x86-64 Base64 decoding intrinsic, the optimized routine the JIT compiler substitutes for the Java code once a method is hot. When the last few characters of a padded input are processed, the routine builds a store mask that can end up covering a whole 64-byte vector. The write lands past the end of the output array, and Java rejects the malformed input only afterwards, by which time neighboring heap objects may already be overwritten. Calif says the corruption can be turned into a Heartbleed-style memory leak, which it calls JavaBleed, or into code execution.
The fix is small. Endor Labs says it “is four lines under JDK-8393212”. The mainline commit, 5177a70b385c in openjdk/jdk, carries the title 8393212: Base64 optimization improvements, which says nothing about security. It adds this to generate_base64_decodeBlock() in src/hotspot/cpu/x86/stubGenerator_x86_64.cpp (a comment, two lines of code and a blank line):
+ // A padded Base64 tail shorter than one complete block is handled by Java.
+ __ cmpl(length, 4);
+ __ jcc(Assembler::less, L_bruteForce);
Per the added comment, a padded tail shorter than one complete block is now left to the Java code instead of being handled by the vector stub. The commit header is dated October 13, 2026 and the co-author line names Albert Mingkun Yang. The OpenJDK 27 update repository carries a backport authored on October 14.
Who is actually exposed
Four conditions have to line up, and the sources agree on most of them.
- A HotSpot JVM from 17.0.6. In the jdk17u source,
stubGenerator_x86_64.cpphas nogenerate_base64_decodeBlockat tagjdk-17.0.5-gaand has one atjdk-17.0.6-ga, which matches the floor Calif, Talos and Red Hat give. 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 OpenJDK advisory’s risk matrix marks 17, 21, 25 and 27, and Red Hat says “Customers running OpenJDK 8 or OpenJDK 11 are also not affected, as these versions do not contain the vulnerable code path.” Oracle’s CVE entry still lists one Java 8 build, 8u511b33-perf, among the affected Oracle Java SE releases. - An x86-64 CPU with AVX-512 VBMI. Calif names the Intel Xeon generations Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD EPYC and Ryzen parts from Zen 4 on. Red Hat’s bulletin lists the cloud families it expects to be most exposed: AWS C6i, M6i, R6i, C7i, M7i and R7a; Google Cloud C3, C3D, C4, C4D, N4, N4D, M3 and M4; and Azure D-series v5 to v7 sizes such as Dsv5 and Dasv6. We did not verify those instance lists.
- Attacker-controlled Base64 reaching the JDK decoder. Red Hat’s examples are HTTP Basic credentials, SAML artifacts, CSRF tokens and MIME encoded words, “frequently decoded before authentication completes.”
- A chain built for the application. Calif’s own words are “there is no universal exploit”, and its write-up adds “Every target needs a bespoke profile.” Its chain also warms the decoder until the JIT selects the vulnerable path, depends on heap layout, and is sensitive to network conditions: “Calif rented another VM in the same region as our target.”
The CVE entry lists none of this
The CVE record, published at 21:32 UTC on November 17, is Oracle’s standard text. It lists the affected Oracle Java SE versions and calls the bug a “Difficult to exploit vulnerability” that “allows unauthenticated attacker with network access via multiple protocols”. We searched the whole record, ignoring case: it contains no mention of Base64, AVX, x86, CPU, VBMI, decoder, intrinsic, HotSpot, JVM or hardware, and the only component it names is “Compiler”. NVD lists its status as “Received” and carries Oracle’s score and CISA’s SSVC block but no configuration data. A scanner working from this entry alone cannot tell an exposed AVX-512 host from an ARM64 host that the vendors say is safe.
One bug, four scores
| Source | Rating | What it assumes |
|---|---|---|
| Calif | CVSS 4.0, 9.2 (Critical) | Pre-authentication code execution in six default-configuration products |
| Red Hat (preliminary) | CVSS 3.0, 9.8; bulletin impact Critical | AV:N/AC:L/PR:N; bulletin RHSB-2026-018, CWE-787 |
| Oracle (the CNA); also NVD and SUSE | CVSS 3.1, 8.1 | AV:N/AC:H/PR:N/UI:N, “Difficult to exploit” |
| Amazon Linux | CVSS 3.1, 8.8, severity Important | AV:N/AC:L/PR:L |
| CISA (SSVC block in the CVE record, 21:53 UTC November 17) | Exploitation none, Automatable no, Technical Impact total | CWE-122, heap-based buffer overflow |
The scores differ on attack complexity and privileges, not on the bug. Oracle’s High attack complexity fits Calif’s “no universal exploit”. Amazon’s description reads “a JVM crash for unauthenticated users or, in certain application environments, code execution by an authenticated low-privileged user”, which matches Calif’s own split between six pre-authentication targets and five authenticated ones (Kafka, Solr, Elasticsearch, Gerrit and Spinnaker). CISA’s alert goes further than either: “Specially crafted Base64 input could crash the JVM and allow for unauthenticated remote code execution.” We saw the same spread for the Brotli decoder flaw, where one CVE carried seven ratings.
Is anyone exploiting it?
No source we read says so. Talos wrote that it “has not observed in-the-wild exploitation of this vulnerability yet” and that “we anticipate probing attempts may pick up soon after public disclosure.” CISA’s SSVC entry says exploitation none, the Known Exploited Vulnerabilities catalog did not list the CVE (its latest release, 2026.11.17, was cut at 17:53 UTC, before the CVE was public, so check it again), and FIRST’s EPSS API returned no score.
VulnCheck, quoted by SecurityWeek, says it has built a full exploit but does not expect the bug to be easy to use: adversaries “are likely to find narrow but functional paths to RCE in specific applications”. Calif says its chains stay private: “The weaponized chains for individual products stay private until everyone has had a chance to upgrade. We will publish them later.” The write-up also says “Request buffering is not a general mitigation”: nginx buffers request bodies by default, which removes the timing some chains need, but the Elasticsearch chain does not depend on it.
On the AI claim, Calif writes: “It took our own Quang Luong an hour to discover the vulnerability, and two more hours to develop the first working exploit. The total token cost was less than $200.” That is the finder’s own account. The same write-up says Calif “was able to accelerate the exploit development by spending initial effort building pipelines and strategies,” and Luong describes ignoring “slops” from Codex for about 20 minutes before a real finding appeared. We cannot check any of it independently.
Which builds are fixed
We read each distributor’s own page between about 10:10 and 10:20 UTC on November 18.
| Build or distributor | Status when we checked | Where we read it |
|---|---|---|
| OpenJDK upstream | Fixed in 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1: the fix comment is present in the source at the four -ga tags and absent at jdk-17.0.21+9, jdk-21.0.13+9, jdk-25.0.5+8 and jdk-27.0.1.0.3-ga. GitHub’s public event feed shows pushes to the release branches for 27.0.1.1 at 17:40 UTC, 21.0.13.1 at 21:20 UTC and 17.0.21.1 at 21:26 UTC on November 17. Endor Labs wrote that as of November 17 the 17, 21 and 25 update repositories “did not yet carry it.” |
raw source at each tag; GitHub events API; Endor Labs |
| Oracle Java SE | Patched in the November 17 update. Oracle’s page lists affected builds 17.0.21.0.3, 21.0.13.0.3, 25.0.5.0.3, 27.0.1.0.3 and 8u511b33-perf. | Oracle advisory |
| Azul Zulu | Listed as fixed in the November 17 release notes, for Zulu 17, 21, 25 and 27. | docs.azul.com |
| Amazon Corretto and Amazon Linux | Fixed on November 17 for java-17 (Amazon Linux 2, 2023 and 2027 Preview), java-21, java-25 and java-27 (2023 and 2027 Preview); Corretto 8 and 11 “Not Affected.” The current Corretto download builds are 17.0.21.1.2, 21.0.13.1.2 and 25.0.5.1.2. | Amazon Linux Security Center; corretto.aws |
| Eclipse Temurin | No fixed release: the Adoptium API’s latest GA builds were still 17.0.21+9, 21.0.13+9 and 25.0.5+8, dated October 22, and the .1 builds existed only as -ea-beta tags. |
api.adoptium.net; GitHub tags |
| Microsoft Build of OpenJDK | Release notes list no build for the November 17 advisory; the newest entries are 25.0.5.0.2 and 21.0.13.0.2, which cite October advisories. | learn.microsoft.com |
| Red Hat | Bulletin RHSB-2026-018 lists RHEL 8, 9 and 10 and Red Hat build of OpenJDK 17, 21 and 25 as affected and says “Red Hat is expediting the release of fixes.” Its CVE page says the upstream fixed version had “not yet been confirmed by Oracle”. | access.redhat.com |
| SUSE | State Pending, rated important; java-17, java-21 and java-25 packages Affected, java-11 and java-1_8_0 Not affected. | suse.com |
| Debian | Unstable fixed (17.0.21.1+2-1, 21.0.13.1+2-1, 25.0.5.1+2-1, 27.0.1.1+4-1); bookworm and trixie security still listed vulnerable. | security-tracker.debian.org |
| Ubuntu | All 25 openjdk-17, openjdk-21, openjdk-25 and openjdk-lts entries (releases plus upstream) read “needs-triage”. | ubuntu.com JSON |
Version labels differ by vendor, so do not compare numbers by eye. Endor Labs puts it bluntly: “Do not infer it from Oracle’s numbering.”
The timeline in the public records
- October 13: the mainline fix is authored, per the commit header. October 14: the 27 update backport is authored.
- October 22: the CVE record is reserved.
- November 16, 19:52 UTC: committer timestamp of the mainline commit. Git metadata does not show when it became publicly visible.
- November 17, 16:53 and 17:14: Red Hat creates and updates its bulletin (the page gives no time zone). The Red Hat statement says the flaw was reported “ahead of Oracle’s planned fix.” SecurityWeek says Calif published “after news of the flaw surfaced ahead of the coordinated disclosure.”
- November 17, 17:40 and 17:44 UTC: a push to the 27.0.1.1 release branch, then the OpenJDK advisory’s “last update” stamp. Calif’s page says “OpenJDK released a security advisory and Oracle released a patch at 1:00pm PT on Nov 17, 2026.” That is 21:00 UTC, more than three hours after the advisory’s own stamp.
- November 17, 21:20 to 22:18 UTC: pushes to the 21.0.13.1 and 17.0.21.1 branches, the CVE publication (21:32), CISA’s SSVC entry (21:53) and the NVD record (22:18).
Two names are credited. Calif’s page says the bug “was found and reported by Calif.io”, while Oracle’s credit statement lists both “Albert Yang” and “Quang Luong of Calif.io” for CVE-2026-114938, and the OpenJDK advisory acknowledges nine names without mapping them to CVEs. No source says whether the Oracle credit and the commit’s co-author line refer to the same person, or which report came first.
What to check on your own servers
These steps are ours, built from the sources above.
Find every runtime, including bundled ones
Run java -version for each application’s own runtime, not just the one on the system path. Red Hat’s bulletin asks for an inventory that includes runtimes “bundled inside containers, product installations, and vendor appliances.” CISA’s alert warns that Java runtimes “may not appear as separate entries in software inventories.” Upstream-based builds are fixed at 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1, and vendor suffixes differ. For products that bundle their own JVM, Red Hat says to “apply the mitigation flag or update via the product vendor’s supported path rather than patching the JVM independently.”
Check whether the hardware can take the vulnerable path
Calif’s Linux command prints whichever of four CPU flags are present, and its page says the output must include all four. We added a count variant so the answer is one number, and tested both on three made-up flag lines, not on real hosts:
# Calif's command: the output must list all four flags
grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' | grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'
# Count variant (ours): prints 4 on hardware with all four flags
grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' | grep -Ec '^(avx512f|avx512bw|avx512vbmi|bmi2)$'
| Made-up flag line | Calif’s command lists | Count variant |
|---|---|---|
| AVX-512 with VBMI | bmi2, avx512f, avx512bw, avx512vbmi | 4 |
| AVX-512 without VBMI | bmi2, avx512f, avx512bw | 3 |
| AVX2 only | bmi2 | 1 |
Inside a virtual machine, /proc/cpuinfo shows what the hypervisor exposes, which is also what the JVM sees. Red Hat’s shorter check, grep -o avx512vbmi /proc/cpuinfo | head -1, printed avx512vbmi on the first made-up line and nothing on the other two.
Do not read UseBASE64Intrinsics as an exposure test
We downloaded Temurin 21.0.13+9, a build in the affected range, onto a Windows 11 desktop with an Intel Core Ultra 5 235, which has AVX2 but no AVX-512. Filtering the output of java -XX:+PrintFlagsFinal -version to the two relevant flags gave (whitespace condensed):
int UseAVX = 2 {ARCH product} {default}
bool UseBASE64Intrinsics = true {product} {default}
The Base64 intrinsic is on by default here even though this CPU cannot take the vulnerable branch. In vm_version_x86.cpp at tag jdk-21.0.13+9 the intrinsic is enabled whenever UseAVX is at least 2, HotSpot clears the AVX-512 VBMI feature bit when UseAVX is below 3, and the decode stub only takes its AVX-512 branch if VM_Version::supports_avx512_vbmi() and VM_Version::supports_avx512bw() both hold. In 17.0.21 the rule is stricter: the intrinsic is enabled only when UseAVX is above 2 and the CPU reports AVX-512 VL and BW. Asking this desktop for -XX:UseAVX=3 printed OpenJDK 64-Bit Server VM warning: UseAVX=3 is not supported on this CPU, setting it to UseAVX=2. Calif’s published trigger needs a VBMI host, so we did not run it, and we cannot say what the same command prints on one.
If you cannot patch yet, pick the flag that fits your fleet
Calif and Red Hat list two interim flags. Red Hat’s bulletin calls -XX:-UseBASE64Intrinsics the preferred option because “it removes the vulnerable code path while leaving other AVX-512 optimizations in place.” The alternative, -XX:UseAVX=2, switches off AVX-512 entirely, and Calif warns that it “will fail to start” on ARM. The source agrees: at jdk-17.0.21+9, UseBASE64Intrinsics is declared in the shared globals.hpp, while UseAVX is declared in globals_x86.hpp and the aarch64 globals file has no such flag. For a fleet that mixes x86-64 and ARM64 nodes, the first flag is the one you can apply everywhere. Talos’s post prints the second as “=XX:UseAVX=2” with an equals sign where the hyphen belongs, so type it rather than pasting it.
Both flags were accepted by Temurin 21.0.13+9 on Windows x64. -XX:-UseBASE64Intrinsics changed the flag to false; -XX:UseAVX=2 changed nothing visible on this host, where UseAVX was already 2 and the flag stayed true; on a VBMI host the source says it removes the vulnerable branch by clearing the VBMI feature bit. If you set them through JAVA_TOOL_OPTIONS, every JVM prints a “Picked up JAVA_TOOL_OPTIONS” line on stderr, which we saw here and which can upset scripts that parse stderr. Red Hat’s own caveat applies: “Treat this as an interim measure only; apply the permanent OpenJDK fix once available and remove the workaround flag.”
Prioritize authentication paths, and do not rely on size limits
Red Hat and CISA both put internet-facing and authentication services first. CISA also lists “restricting and validating Base64 input at the application boundary, enforcing input-size limits, and limiting unauthenticated access” as compensating controls. Calif says the corruption can be triggered by a string of about 65 characters, so an input-size limit will not stop it. Strict validation before the decoder could help where an application can do it, but we did not test any validator against the bug.
What we could not confirm
- Calif’s per-product code-execution results. The chains are withheld, and we found no notice on the Jenkins advisories page or the Keycloak security page when we looked. We did not check Atlassian, JFrog or Ping Identity.
- Any exploitation. No source reports it, and the KEV and EPSS entries above are a snapshot, not a forecast.
- Calif’s account of how long discovery took and what it cost, which comes only from Calif.
- Behavior on a VBMI-capable host. This machine cannot take the vulnerable branch, so nothing here reproduces the bug.
Sources
- Calif, Java4Shell write-up and OpenJDK Vulnerability Advisory 2026/11/17
- Oracle November 2026 Critical Security Patch Update and the NVD record
- CISA alert, Cisco Talos, Endor Labs
- SecurityWeek and BleepingComputer
- Red Hat bulletin RHSB-2026-018 and CVE page
- Amazon Linux Security Center, SUSE, Debian, Ubuntu
- Azul release notes, Microsoft Build of OpenJDK release notes, Adoptium API








No Comment! Be the first one.