TL;DR - Java4Shell is CVE-2026-114938, a memory corruption bug in OpenJDK's AVX-512 Base64 decoder, disclosed on 17 November 2026 with Oracle patches out the same afternoon. Calif rates it CVSS 4.0 9.2 and Oracle rates it CVSS 3.1 8.1. Unauthenticated remote code execution was demonstrated in default configurations of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate. Four conditions must all hold: OpenJDK 17.0.6 or later on the 17, 21, 25 or 27 lines, an x86-64 CPU with AVX-512 VBMI, the AVX-512 Base64 intrinsic enabled, and attacker-influenced Base64 reaching java.util.Base64. No in-the-wild exploitation has been reported. What you need to do: run the CPU check first, inventory every bundled JRE, patch internet-facing Java tonight, and log the de-scoped hosts with a date and a reason.
By The Numbers
| Figure | Value |
|---|---|
| CVE | CVE-2026-114938, named Java4Shell by the finders |
| Severity, as rated by Calif | CVSS 4.0, 9.2 critical |
| Severity, as rated by Oracle | CVSS 3.1, 8.1 |
| Disclosure and patch release | 17 November 2026, Oracle patches at 1pm PT |
| Conditions that must all hold for exploitability | four |
| Affected OpenJDK lines | 17, 21, 25 and 27, version 17.0.6 or later |
| Products with demonstrated unauthenticated RCE in default config | six: Keycloak, Jira, Confluence, Artifactory, Jenkins, PingFederate |
| When the bug was first reported in the wild | 9 June 2026, in an Adoptium issue |
Every time a Java vulnerability makes the front page, I get the same message from at least two clients: "we don't run Java, so we're fine." Then I open their Hyper-V console and find four appliances running an embedded JRE, a backup server with its own runtime, a build agent nobody has logged into since a contractor left, and a Jira instance that the marketing team swears is just a website.
Java4Shell is that conversation with a deadline attached. It was disclosed on 17 November, patches went out the same afternoon Pacific time, and it has already been demonstrated as unauthenticated remote code execution in the default configuration of six products that small businesses actually run. It will still be open in a lot of estates next month for one reason: the affected thing is invisible, because it is a runtime inside something else.
I am not going to write the same news piece everyone else wrote. This is the triage order I would use, starting with the four conditions that have to hold at the same time, because getting through them fast is how you find out whether you have a real problem or a paperwork problem.
Let me walk you through the four conditions, where the hidden runtimes live, the patch order I would follow, and the honest limits of what a WAF or a firewall buys you here.
What the bug actually is
The short version, because the exploit mechanics are less useful to an admin than the conditions.
OpenJDK ships a hand-optimised Base64 decoder that uses AVX-512 instructions on x86-64 chips that support them. It exists because Base64 decoding is hot in a lot of workloads and doing it in vector registers is much faster than doing it byte by byte. Java4Shell is a memory corruption bug in that decoder. A malformed padded tail in the input makes the decoder write past the end of the output array.
What happens next depends on how much effort the attacker puts in. At the low end you get a JVM crash, which is a denial of service and nothing more. In the middle you get a Heartbleed-style heap read, where memory past the array is returned to whoever sent the input. At the high end you get unauthenticated remote code execution, and the researchers behind the disclosure demonstrated exactly that against six products running default configurations.
That ladder is worth understanding, because it tells you what a half-patched estate looks like. You do not choose which rung your attacker stops at. The same input shape that crashes a service is the one that leads to the interesting outcomes.
The six products with demonstrated RCE matter for triage order: Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate. Every one is a server in front of a login form or a build pipeline, decoding Base64 from untrusted input as part of normal operation.
The four conditions, and why they are the whole exercise
Calif's advisory lists four conditions that must all be true for exploitation. This is the most useful thing about the disclosure, because it means most estates can be de-scoped in an afternoon and the ones that cannot need to move today.
Condition one: OpenJDK 17.0.6 or later, on the 17, 21, 25 or 27 lines. Standard Java 8 and Java 11 are not affected, because the vulnerable decoder is not present in the form that matters. There is one listed exception, an Oracle Java 8 performance build labelled 8u511b33-perf, which is the kind of thing almost nobody is running deliberately and quite a lot of people are running without knowing.
Condition two: an x86-64 CPU with AVX-512 VBMI. In Intel terms that means Ice Lake and later. In AMD terms that means Zen 4 and later, so EPYC Genoa and Bergamo, not Rome or Milan. ARM64 is not affected at all.
Condition two does the most work in most small environments. A fleet of virtual machines sitting on five-year-old Xeon Scalable or second and third generation EPYC hosts fails it, and those hosts are extremely common in SMB infrastructure. Before you spend a night patching, run one command on one host and find out which side of the line you are on.
Condition three: the AVX-512 Base64 intrinsic enabled in the JVM. This depends on the build and on any JVM flags in use. Check the vendor advisory for the specific build you are running rather than assuming a default, because this is the condition where folklore is most likely to be wrong.
Condition four: attacker-influenced Base64 reaching java.util.Base64. This is the interesting one for defenders, because it is a property of the application rather than the runtime. Lots of applications decode Base64. Fewer decode Base64 that an unauthenticated stranger controls. The six demonstrated products all do.
All four or nothing. That is genuinely good news for a lot of people reading this, and it is also the reason to check rather than assume, because the conditions are the sort of thing that changes quietly when somebody adds a host to the cluster or rewrites a deployment to a newer base image.
Where the hidden runtimes are
If you are confident you do not run Java, this is the list to work through before you stay confident.
Web applications you think of as products rather than Java. Jira, Confluence, Artifactory and Jenkins are Java applications. So is Keycloak, which a lot of identity projects deploy and forget. If your "website" has a /jira path or a .war file in its deployment, it is Java.
Build servers and CI runners. Every build agent may carry its own JDK, at whatever version the last person needed. Build servers also tend to be reachable from the internet in organisations that would never expose a database, because someone needed webhooks to work.
Backup clients and management agents. Backup software, monitoring agents and endpoint management tools often ship a bundled runtime so the vendor does not depend on yours. That runtime sits on your host and does not appear in your application inventory.
Virtual appliances. Anything shipped as an OVA or a vendor-managed VM may well have a JRE inside its management plane. Network management, log aggregation, ITSM connectors and document management are the usual suspects.
Small tools that became necessary. A utility someone installed to convert reports, an integration bridge between two systems, a scheduled job running since 2023. Each one is a runtime, and each one has an owner who has left.
Container images. Base images get rebuilt from cached layers. If a team has not rebuilt since June, the image still carries a vulnerable runtime, and your scanner will say so if you point it at the registry rather than the host.
The triage order I would follow
1. Run the CPU check first, on every host and every cluster node. This is free and it de-scopes fastest. On a Linux host, look for the AVX-512 VBMI flag in the CPU flags. On a virtual fleet, this is a question about the physical hosts and their CPU model, so check that before you check anything else. If the whole estate is on hardware that fails condition two, you have a documentable finding rather than a patch emergency, and you can still patch on your normal cycle.
2. Inventory the runtimes you did not know about. Every JDK and JRE, on every host, in every image, including the bundled ones. This is the slow step and it is the one that pays off twice, because a runtime inventory is the thing you should have had anyway.
3. Check versions against condition one. Anything 17.0.6 or later on the 17, 21, 25 or 27 lines is in scope. Java 8 and 11 come off the list, with the Oracle performance build exception noted above.
4. Rank what is internet-facing. The six products with demonstrated RCE are all things that sit behind a login or a webhook, and the pattern generalises. Public-facing Java that decodes Base64 from untrusted input patches first, tonight if you can. Internal admin consoles patch second. Everything else goes in the normal cycle.
5. Patch in the vendors' own version numbers, not the line number. Oracle shipped 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. Amazon Corretto shipped 17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2 and 27.0.1.1.4. Those are the specific builds to match, and distributors differ, so read the advisory for the runtime you actually run rather than the one you think you run.
6. Deal with the distribution you are stuck on. Temurin had no confirmed fix at the time of disclosure. That is not a reason to relax, it is a reason to know which of your applications run on Temurin and what your fallback is: move to a patched distribution, apply the vendor's mitigation, or restrict reachability to the service until a build lands. Writer's note: check the current status of your distribution's advisory rather than trusting this paragraph, because this is the part of the story most likely to have changed by the time you read it.
7. Record the de-scoped hosts as findings. "Not affected, CPU lacks AVX-512 VBMI, checked 22 November" is a real result and belongs in your vulnerability register. It has an expiry date, because the next hardware refresh deletes the exemption.
The patch-window context
This disclosure is the second from the same research group in a month, and the earlier one was PHP2Shell on 21 October. Same team, same naming convention, different runtime. That sibling post is worth reading if you want the pattern, because the two triages overlap in the way that matters: both are cases where the vulnerable component is embedded in something else and the asset inventory is the actual work.
The difference is timing pressure. PHP2Shell was about how fast you can move a patch through a change process. Java4Shell is about whether you know where the application even is, which is a slower and more structural problem.
No exploitation in the wild has been reported, and as of 18 November the vulnerability was not in CISA's known exploited vulnerabilities catalogue. CISA did issue an alert on 17 November urging immediate patching, which is a signal about expected weaponisation rather than evidence of current attacks.
The disclosure timeline worth one paragraph
The history here is genuinely odd and it is the part security people will be arguing about for a year.
The bug was first found in the wild on 9 June 2026, in an Adoptium issue. A fix was proposed on 9 September, as openjdk/jdk#32781. It sat there, as these things do. Then on 24 October a researcher working with Codex rediscovered the same bug independently, and the rediscovery is what moved it to a coordinated patch on 17 November.
Read those three dates together. The defect was reported in June, a fix existed from September, and what finally produced a patch release was an AI-assisted researcher arriving at the same bug from a different direction. I do not know why the September fix did not ship and I am not going to guess. What I can say is that "the fix exists" and "the fix is deployed" were different states for over four months, and Java4Shell is the name of the gap.
Why this could be good
There are two upside readings here and I think both are honest.
The first is that four required conditions is a lot of conditions. Plenty of organisations will read the advisory, run the checks and find they are not exposed in any meaningful way. That is a good outcome, and a better use of an afternoon than the panic that follows a wormable vulnerability with no prerequisites.
The second is that the four-condition disclosure style is what good research looks like from an operator's point of view. Giving admins the conditions means they can triage without becoming exploit developers. Compare it with the disclosure that says "critical, patch everything" and leaves you working out what "everything" means at 9pm.
And a third upside: this is a useful forcing function for a runtime inventory. Most organisations cannot answer "how many JVMs do we run and where" today. After this week, a lot of them can.
Where this is genuinely concerning
Temurin had no confirmed fix at disclosure. Adoptium is one of the most widely used OpenJDK distributions in the world, and a lot of container images and CI pipelines default to it. If a chunk of your estate runs on it, your patch window is not in your control, and the honest answer is to know which systems are in that bucket before you need to.
The four conditions are true until they are not. Hardware gets replaced, clusters migrate to newer hosts, deployments move to a base image with a different runtime. Each change can flip a de-scoped host into scope, and none of them raise a security ticket. The de-scope is a snapshot.
The CVSS disagreement is large. Calif rates it 9.2 under CVSS 4.0 and Oracle rates it 8.1 under CVSS 3.1. The scoring systems are not directly comparable, so quoting "9.2" in a risk register and moving on is doing the arithmetic wrong. The gap reflects different assumptions about exploitability.
The AI-discovery precedent cuts both ways. A bug reported in June, fixed on paper in September, patched in November after an AI-assisted researcher found it again. That is one more data point that AI tooling will surface a lot of these, and a reminder that the bottleneck sits in the path from a proposed fix to a shipped build.
The "we don't run Java" reflex is still the biggest risk. The demonstrated products are all things somebody thinks of as a tool rather than a server. If your inventory is built from the application list rather than the host list, the affected product is not on it.
What to do right now
- Run the CPU check on every physical host and every cluster node. Look for AVX-512 VBMI support. On most SMB fleets this alone resolves the question for the majority of virtual machines, and it takes about ten minutes.
- List every Java runtime on every host, including bundled ones. Agents, backup clients, appliance management planes, CI runners, container images. If your inventory is application-based, rebuild it host-based for this exercise.
- Check versions against 17.0.6 and later on 17, 21, 25 and 27. Cross off Java 8 and 11, with the Oracle 8u511b33-perf build as the listed exception.
- Patch internet-facing Java first. Start with Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate if you run any of them, then work outwards to anything else that decodes Base64 from untrusted input.
- Match the exact fixed builds. Oracle 17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1. Corretto 17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2, 27.0.1.1.4. Check the advisory for your own distribution, especially if it is Temurin.
- Write the mitigations down for anything you cannot patch tonight. Restrict reachability at the network layer, put the service behind authentication you control, and check the vendor's own mitigation guidance. Treat edge filtering of malformed Base64 as a speed bump rather than a fix.
- Log the de-scoped hosts with a date and a reason. Note the CPU model, the runtime version and the check you ran. It becomes evidence later, and it becomes a review item when the hardware changes.
- Add runtime inventory to your normal cycle. The next one of these will have different conditions and the same problem, which is that you do not know where the runtime is until somebody asks.
Key Takeaways
- Java4Shell is CVE-2026-114938, a memory corruption bug in OpenJDK's AVX-512 Base64 decoder, disclosed on 17 November 2026 with patches released the same day.
- Four conditions must all be true: OpenJDK 17.0.6 or later on the 17, 21, 25 or 27 lines, an x86-64 CPU with AVX-512 VBMI, the AVX-512 Base64 intrinsic enabled, and attacker-influenced Base64 reaching
java.util.Base64. - Unauthenticated remote code execution was demonstrated in default configurations of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate.
- The CPU requirement de-scopes the most estates, because Intel Ice Lake and AMD Zen 4 are the floor and a lot of SMB virtual fleets run on older silicon.
- The runtime is usually bundled inside something else. Build agents, backup clients, appliance management planes and container images are where the invisible JVMs live.
- Temurin had no confirmed fix at disclosure, so knowing which of your systems depend on it matters more than a general instruction to patch.
- No in-the-wild exploitation has been reported and the vulnerability was not in CISA's KEV catalogue as of 18 November, but CISA issued an alert on 17 November urging immediate patching.
FAQ
Am I affected if I run Java 8 or Java 11? No, not on the standard builds. The affected decoder is in OpenJDK 17.0.6 and later on the 17, 21, 25 and 27 lines. Oracle lists one Java 8 performance build, 8u511b33-perf, as an exception, so check whether you are running a non-standard build before you dismiss it.
Is Apple Silicon or ARM affected? No. The bug is in the AVX-512 code path, which only exists on x86-64. ARM64 hosts are not affected by this condition.
How do I check whether my CPU supports AVX-512 VBMI? On Linux, inspect the CPU flags and look for the VBMI flag, and confirm the physical host model rather than the virtual CPU if you are on a hypervisor. Intel Ice Lake and later, and AMD Zen 4 and later, are the generations that carry it.
Do I need to patch if the service is internal only? Check all four conditions rather than deciding on reachability alone. Internal-only reduces the likelihood of someone sending crafted input, and it does not reduce the impact if a less-privileged user or a compromised workstation can reach the service.
Which fixed builds should I look for? Oracle shipped 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. Amazon Corretto shipped 17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2 and 27.0.1.1.4. Other distributions have their own build numbers, so use the number from your own vendor's advisory.
Will a WAF protect me until I patch? Treat it as a speed bump. Filtering malformed Base64 at the edge can reduce exposure, and the demonstrated exploits run against default application configurations that a generic rule set was not written for. Network restrictions and moving the service behind authentication you control are stronger interim controls.
Is this tied to PHP2Shell? Same research group, same naming convention, disclosed on 21 October 2026 and covered here at the time. Different runtime and different mechanism. The shared lesson is that the vulnerable component is usually embedded in something else.
Was this exploited in the wild? No exploitation has been reported. It was not in CISA's known exploited vulnerabilities catalogue as of 18 November, and CISA issued an alert on 17 November urging immediate patching.
My Take
The part of this story I keep returning to is the timeline, because it says something uncomfortable about how the industry works now.
A defect was found in June. A fix was proposed in September. Nothing shipped. A researcher with an AI coding assistant found the same bug again in late October, and the patch landed on 17 November. So the thing that finally fixed it was not more analysis, and it was not a vendor deciding to prioritise. It was rediscovery, which is a strange thing to rely on.
The implication I would draw is that discovery is becoming cheap and delivery is not. If AI tooling makes finding bugs like this routine, and the evidence of the last two months suggests it might, then the constraint moves to the part nobody wants to talk about: the vendor's path from a proposed patch to a shipped build, and your own path from a shipped build to an installed one. Those are process problems dressed up as security problems, and they are where the next four months will be spent.
For this week, though, the job is smaller and more concrete. Run the CPU check. Build the runtime list. Patch the internet-facing Java tonight. Write down the de-scoped hosts with the reason and the date, because the reason will expire before the risk does.
The organisations that come out of this well will not be the ones that patched fastest. They will be the ones that can finally answer how many Java runtimes they own.
Mathew Clark Founder, SecureInSeconds Currently: finding four JVMs on a network that "doesn't run Java" Drafted with AI assistance - read, verified, and fixed by a human.
Further Reading
- Calif advisory for Java4Shell, CVE-2026-114938 - the four conditions, the demonstrated products and the severity reasoning
- Oracle Critical Patch Update, 17 November 2026 - the fixed build numbers for the affected OpenJDK lines
- CISA alert on Java4Shell - the agency's guidance and current known exploited vulnerabilities catalogue
- openjdk/jdk#32781 - the September fix that did not ship in time
- Armorcode's analysis of Java4Shell exposure and Pentera's write-up - two independent reads on exploitability and detection
- PHP2Shell and the patch window - the sibling disclosure from the same researchers, and the patch-window problem it raised
- Hashicorp Vault RCE and the OpenBao patch - what an embedded runtime under a login form looks like when it breaks
- Atlassian CVE-2026-21589 pre-auth file read - the last time Jira and Confluence were the reason to patch this month



