Java4Shell (CVE-2026-114938) and how IONIX can help
This is a rolling post. Last updated: 17 Nov 18:40 UTC
In this article
TL;DR
CVE-2026-114938 (“Java4Shell”) is a reported memory-corruption flaw in the Java runtime that can reportedly be triggered over the network without authentication. Early reports describe it as remote code execution (RCE). So far the details are not confirmed, and no exploitation or public exploit has been reported.
Our own early analysis shows the flaw is narrower than “all of Java”. It affects only some Java 17, 21 and 25 servers running on certain CPUs, and only when an attacker can send crafted Base64 data to the application. Most Java assets are likely not exposed.
What to do now: find your Java 17, 21 and 25 systems, check them with the commands below, and apply the vendor fix once it is confirmed.
What to expect from IONIX, and when
- Already in the platform: assets that run Java or a known Java-based product (such as Tomcat, Spring Boot, Jira, Confluence or WebLogic), identified from data we already collect, are flagged in the Threat Center post now. This list is a starting point for triage, not a confirmed list of vulnerable assets. The Java version and the server’s CPU cannot always be determined from the outside, so it includes assets that are not affected.
- Narrowed list of potentially affected assets: expected by the end of day on 17 Nov (UTC). It will focus on products that bundle or require an affected Java version (in cases where such versions are visible from the outside). It will still be a list of candidates to focus on first / to prioritize. You may still need to eventually inspect all Java assets for their versions.
- Safe exploitability test (confirmed findings): not expected. Triggering the flaw writes outside allocated memory and can crash the target Java process, so we will not run it against customer assets.
- WAF rules: available now. IONIX has published interim WAF rules for Cloudflare, AWS WAF and Akamai, tested in front of a vulnerable and a patched Java lab. They partially mitigate the risk and don’t replace patching. See “Interim WAF rules” below.
- Updates to this post: when the technical details and the official advisory are published, and when fixed builds are confirmed.
Key dates
- 17 Nov: Oracle’s November 2026 Critical Patch Update is scheduled for release. Oracle has not said whether it covers this CVE.
- 17 Nov, 21:00 UTC: the researchers who reported the flaw (Calif) say they will publish technical details after Oracle releases its patch. This is the expected end of the embargo, based on their public post, and it may change.
- The CVE record is not yet published in the NVD. We will update this post when it is.
Affected software
Based on IONIX research of the OpenJDK source, not yet confirmed by Oracle or OpenJDK:
- Affected: HotSpot-based Java 17.0.6 and later, Java 21, and Java 25 (OpenJDK and builds based on it, such as Temurin, Corretto, Zulu and Red Hat builds, and Oracle Java SE).
- Not affected: Java 17.0.5 and earlier, and Java 11. Java 8 was not found to contain the affected code in the reports we reviewed.
- Not affected: JVMs that do not use HotSpot, such as OpenJ9 and GraalVM native images.
Reports so far mention only Java 17. Oracle’s pre-release notice lists several Java versions without tying them to this CVE. We will correct this section when the official advisory is out.
Who is actually exposed
Based on initial (unconfirmed) research, a server is currently believed to be exposed only if all of the following are true:
- It runs an affected Java version (above).
- It runs on an x86-64 CPU with AVX-512 VBMI support (Intel Ice Lake or newer, AMD Zen 4 or newer). ARM servers and older Intel servers are not affected.
- The application decodes Base64 that an attacker controls, before any login, using the standard Java Base64 decoder. Common examples are HTTP authorization headers and tokens.
Many Java servers will not meet every condition. We therefore expect the number of truly exposed servers to be much smaller than the number of Java servers.
Technical details
- Component: the Java runtime (reported as a memory-corruption flaw in a Base64 decoding routine).
- Attack: over the network, no authentication, no user interaction.
- Impact: reported as RCE. In our testing, the flaw writes data outside the memory it should use. We have demonstrated a crash via the related code path in a lab environment. The most likely practical impact is a crash or instability, and RCE remains unconfirmed.
- Exploitation status: no exploitation and no public exploit reported as of the last update.
Severity
CVSS 4.0: 9.2 (Critical), as reported before official disclosure. The final score will follow the official advisory.
How to check a Java server yourself
Run these on the host. They are safe and read-only:
java -version
lscpu | grep -o avx512_vbmi
java -XX:+PrintFlagsFinal -version | grep -E "UseAVX|UseBASE64Intrinsics"
Do not test with crafted input against production systems, because it can corrupt memory and crash the application.
Mitigation and recommended actions
- Inventory your Java 17, 21 and 25 systems, starting with the assets IONIX lists for you. Products often bundle their own Java (for example ColdFusion, Adobe Experience Manager and Zimbra), so check vendor advisories too.
- Apply the fix once it is confirmed. As of this update no fixed build is confirmed, and Oracle’s notice lists current builds as affected. Check that the release notes name CVE-2026-114938 before assuming a build is fixed. Rebuild container images and restart services after updating.
- Interim workaround (IONIX research, not vendor endorsed): start the JVM with
-XX:-UseBASE64Intrinsics. In our testing on Java 17.0.21 this stopped the out-of-bounds write. It requires a restart and may slow Base64 decoding slightly. Setting-XX:UseAVX=2had the same effect. - Reduce exposure: remove unneeded internet-facing Java endpoints and restrict access with a VPN or allowlist.
- WAF rules: a WAF can help but is only a partial control, because the crafted data can appear in many places, such as headers, tokens and request bodies. See the interim rules below.
Interim WAF rules
IONIX Research Labs, not vendor endorsed. The rules block the input shape that triggers the flaw: one or more whole blocks of 64 Base64 characters, followed by =, ==, x=, x== or xx=. Correctly padded Base64 never has this shape, so normal traffic is not blocked.
We tested these rules on AWS WAF, Cloudflare and Akamai, in front of a vulnerable and a patched Java lab. Each WAF has gaps, listed below. A WAF rule reduces exposure but does not replace patching.
Regexes. Apply them with no decoding transformation:
- General, for header, cookie, query and body values:
(?:^|[^A-Za-z0-9+/_%-])(?:(?:[A-Za-z0-9+/_-]|%[0-9A-Fa-f]{2}){64})+(?:(?:[A-Za-z0-9+/_-]|%[0-9A-Fa-f]{2}){0,2}(?:=|%3[Dd])|(?:[A-Za-z0-9+/_-]|%[0-9A-Fa-f]{2})?(?:=|%3[Dd]){2})(?:$|[^A-Za-z0-9+/_=%-])
- Path, for the raw request path:
^[^?#]*/(?:(?:[A-Za-z0-9+_-]|%[0-9A-Fa-f]{2}){64})+(?:(?:[A-Za-z0-9+_-]|%[0-9A-Fa-f]{2}){0,2}(?:=|%3[Dd])|(?:[A-Za-z0-9+_-]|%[0-9A-Fa-f]{2})?(?:=|%3[Dd]){2})(?:$|[/?#;])
- Parsed values, for values the WAF has already decoded (AWS
JsonBody):
(?:^|[^A-Za-z0-9+/_-])(?:[A-Za-z0-9+/_-]{64})+(?:[A-Za-z0-9+/_-]{0,2}=|[A-Za-z0-9+/_-]?==)(?:$|[^A-Za-z0-9+/_=-])
Gaps that apply to every WAF
- Anything past the WAF’s inspection limit isn’t seen. If the application trims whitespace before decoding, an attacker can pad a short trigger past that limit.
- Compressed (
gzip) request bodies, multipart file contents, XML entities (=), and text split over repeated headers or parameters that the application joins.
Cloudflare (tested)
Custom rule, action Block. matches needs a Business or Enterprise plan, and http.request.body.raw needs Enterprise.
(any(http.request.headers.values[*] matches r"(?:^|[^A-Za-z0-9+/_%-])(?:(?:[A-Za-z0-9+/_-]|%(?:25)?[0-9A-Fa-f]{2}){64})+(?:(?:[A-Za-z0-9+/_-]|%(?:25)?[0-9A-Fa-f]{2}){0,2}(?:=|%(?:25)?3[Dd])|(?:[A-Za-z0-9+/_-]|%(?:25)?[0-9A-Fa-f]{2})?(?:=|%(?:25)?3[Dd]){2})(?:$|[^A-Za-z0-9+/_=%-])") or http.request.uri.query matches r"(?:^|[^A-Za-z0-9+/_%-])(?:(?:[A-Za-z0-9+/_-]|%(?:25)?[0-9A-Fa-f]{2}){64})+(?:(?:[A-Za-z0-9+/_-]|%(?:25)?[0-9A-Fa-f]{2}){0,2}(?:=|%(?:25)?3[Dd])|(?:[A-Za-z0-9+/_-]|%(?:25)?[0-9A-Fa-f]{2})?(?:=|%(?:25)?3[Dd]){2})(?:$|[^A-Za-z0-9+/_=%-])" or raw.http.request.uri.path matches r"^[^?#]*/(?:(?:[A-Za-z0-9+_-]|%(?:25)?[0-9A-Fa-f]{2}){64})+(?:(?:[A-Za-z0-9+_-]|%(?:25)?[0-9A-Fa-f]{2}){0,2}(?:=|%(?:25)?3[Dd])|(?:[A-Za-z0-9+_-]|%(?:25)?[0-9A-Fa-f]{2})?(?:=|%(?:25)?3[Dd]){2})(?:$|[/?#;])" or http.request.body.raw matches r"(?:^|[^A-Za-z0-9+/_%\\-]|\\[bfnrt\x22]|\\\\|\\u00(?:[01][0-9A-Fa-f]|20))(?:(?:[A-Za-z0-9+/_-]|%(?:25)?[0-9A-Fa-f]{2}|\\u00[0-9A-Fa-f]{2}|\\/){64})+(?:(?:[A-Za-z0-9+/_-]|%(?:25)?[0-9A-Fa-f]{2}|\\u00[0-9A-Fa-f]{2}|\\/){0,2}(?:=|%(?:25)?3[Dd]|\\u003[Dd])|(?:[A-Za-z0-9+/_-]|%(?:25)?[0-9A-Fa-f]{2}|\\u00[0-9A-Fa-f]{2}|\\/)?(?:=|%(?:25)?3[Dd]|\\u003[Dd]){2})(?:$|[^A-Za-z0-9+/_=%\\-]|\\[bfnrt\x22]|\\\\|\\u00(?:[01][0-9A-Fa-f]|20))" or (http.request.body.truncated and http.request.body.raw matches r"(?:[A-Za-z0-9+/_-]|%[0-9A-Fa-f]{2}){64,}(?:%[0-9A-Fa-f]?)?$"))
Known gaps:
- Cloudflare inspects only the first 128 KB of a body. Trigger text placed after it passes.
- The last clause (
http.request.body.truncated) blocks any body cut off inside Base64 text, including valid Base64 bodies over 128 KB, such as JSON file uploads. Drop it, or scope it to endpoints that don’t take uploads.
AWS WAF (tested, on an Application Load Balancer)
Eight regex-match rules, action Block, OversizeHandling: CONTINUE:
| Field to match | Regex | Transformation |
|---|---|---|
Headers (exclude cookie), values | General | NONE |
Cookies (all), values | General | NONE |
QueryString | General | NONE |
Body | General | NONE and JS_DECODE (two rules) |
JsonBody (all, values, EVALUATE_AS_STRING) | Parsed values | NONE and JS_DECODE (two rules) |
UriPath | Path | NONE |
Known gaps:
- On a load balancer, AWS WAF inspects only the first 8 KB of headers, cookies and body. A trigger longer than 8 KB, or a short one placed after 8 KB of other headers or of whitespace in the body, passes.
OversizeHandling: MATCHcloses this, but blocks every request with more than 8 KB of headers, cookies or body, valid or not. - A cookie with a 64-character name and an empty value passes.
Akamai (tested)
App & API Protector custom rules have no regex operator, so build the rule in Property Manager: extract each header, cookie, the query string and the path into a variable with setVariable, test it with the regularExpression criterion (General or Path regex), and denyAccess on a match.
Known gaps:
- The request body isn’t covered, because Property Manager can’t read it.
- Only the headers and cookies you name are checked.
- Trigger text of more than about 8 KB in
Authorization, the query or the path passed.
What to watch for
The more likely sign of an attack is instability, not a shell. Watch for unexpected Java process crashes and hs_err_pid crash logs. A spike of “Input byte array has incorrect ending byte” errors on endpoints that do not require login can also be a hint, although ordinary malformed requests cause the same error. Also watch for unusual process launches or outbound connections from Java services.
What to watch for
The more likely sign of an attack is instability, not a shell. Watch for unexpected Java process crashes and hs_err_pid crash logs. A spike of “Input byte array has incorrect ending byte” errors on endpoints that do not require login can also be a hint, although ordinary malformed requests cause the same error. Also watch for unusual process launches or outbound connections from Java services.
