Defensive tooling and a technical assessment of the HotSpot Base64 decoding memory-corruption issue tracked in our campaign as Java4Shell / CVE-2026-114938.
Calif now reports application-specific remote code execution, including unauthenticated attacks. Prompt remediation is warranted. Exploitation of this path requires an affected JVM, x86-64 with AVX-512 VBMI and BW, an enabled intrinsic, and attacker-controlled data reaching the relevant decoder. Update affected runtimes; prioritize exposed applications, reported RCE targets and custom code with reused destination buffers. There is no evidence here supporting a blanket shutdown of Java infrastructure.
This repository publishes runtime inventory, source-review tools and passive ModSecurity indicators, not an exploit or an active remote vulnerability probe. Our campaign reproduced memory corruption and laboratory DoS; we have not independently reproduced Calif's product RCE chains. See evidence and limits.
Calif's technical disclosure reports RCE in default application configurations, subject to the vulnerable JVM and CPU prerequisites:
| Reported access requirement | Products named by Calif |
|---|---|
| Unauthenticated RCE | Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins, PingFederate |
| Authenticated RCE | Apache Kafka Broker, Apache Solr, Elasticsearch, Gerrit, Netflix Spinnaker |
Calif also reports RCE in example Spring Boot and Apache Struts applications. Its publication includes technical analysis, a demonstration and a bounded local PoC; product-specific exploit chains are withheld at this review snapshot. There is no universal exploit: application-specific chains are required. These are Calif's reported results, distinct from our own tests below.
Our pre-authentication Elasticsearch decoder finding does not establish pre-auth RCE; Calif places Elasticsearch in its authenticated category. Tomcat's absence from their list does not establish safety. Neither the disclosure nor our lab results alone establish exploitation in the wild.
The CVE is now listed in the official OpenJDK advisory and Oracle security update. See vendor response and release status: an open PR, a mitigation and a released product with a fixed runtime are different states.
Do not wait for a crash before patching a reachable affected runtime. The campaign found JDK Base64 call paths in the products below. They use freshly allocated output arrays: the memory write can be externally silent, and an ordinary authentication error or healthy service does not establish safety. "Quiet sink" names this allocation pattern; it does not guarantee a quiet outcome.
This is an update-prioritization list, conditional on the runtime prerequisites and the named feature being reachable. It does not mean every release or deployment of each product is exploitable. The defect is in the JVM: a product update is sufficient only if it actually replaces/fixes the runtime used by the service.
| Product / stack | Reviewed route to JDK Base64 | Evidence and scope |
|---|---|---|
| Apache Tomcat / TomEE | Basic authentication; Tomcat also has HTTP/2 decoding paths | Tomcat switch: 9.0.89 / 10.1.23, also 11.x; TomEE 10.x/11 inherits newer Tomcat. Tomcat lab + source; TomEE dependency review |
| Spring Boot applications using Spring Security | Servlet HTTP Basic credential conversion before authentication | Security 5.8/6.x and reviewed main; Boot 3.2.5/Security 6.2.x lab. Not every Boot application enables HTTP Basic |
| Elasticsearch | Basic / API-key decoding; separate SAML IdP signature path | Basic lab on 8.11.4; other paths source-reviewed. Check the bundled or explicitly configured JDK |
| Jetty, Jenkins, Solr, Zimbra | Basic/authentication handling in product filters or the embedded container | Source-reviewed paths; Solr Basic must be configured, Zimbra entry concerns its Jetty path |
| Jira / Confluence Data Center | Seraph Basic credential decoding | Bytecode-reviewed Seraph 4.0.4/5.0.7/6.0.2; deployed JVM/version remains decisive |
| Neo4j, SonarQube | HTTP authentication parsing | Source-reviewed paths; version and feature details in the matrix |
| Keycloak | SAML artifact decoding | Campaign source record for the SAML artifact path; do not generalize the separate JWT-path negative result |
| Openfire, Kafka, Apache James | SASL initial response, SCRAM fields, or mail authentication decoding | Source-reviewed protocol paths; authentication mechanism and protocol state matter |
| NiFi, ActiveMQ Classic | Encoded proxied-entity handling; ActiveMQ web/REST authentication | Source-reviewed paths; not a claim about every messaging protocol |
| JDK HTTP server / Jolokia JVM agent | JDK BasicAuthenticator |
Source-reviewed JDK path and Jolokia agent subclass; other Jolokia packaging differs |
| Open Liberty | HTTP/2 HTTP2-Settings decoding |
Source-reviewed h2c path; this is distinct from the reviewed Basic-auth codec |
Library-level entry points also widen the update scope: protobuf-java-util
4.33.0 JSON bytes parsing, SnakeYAML !!binary, pac4j, Spring
Base64Utils, auth0 java-jwt 3.15.0+, fusionauth-jwt, and the campaign's
reviewed Quarkus, Micronaut, RESTEasy, Helidon and fastjson2 paths. Applications must
actually pass untrusted data through the named feature. Older protobuf and JWT
library versions can use different codecs; package presence alone is insufficient.
Product-level runs: the documented Elasticsearch, Tomcat and Spring Security runs exercised decoding paths without an observed JVM crash. Other entries: source, bytecode or dependency evidence, not a reproduced product DoS. No reused destination-buffer call site was established in the audited product paths.
See the full product/library matrix for named call sites, version qualifications, negative paths and unresolved candidates. If a listed path is reachable on an affected runtime, schedule the runtime update even when normal monitoring is silent; use the verified interim mitigation if rollout must wait. Review custom reused buffers urgently, without downgrading fresh-array paths or the applications for which Calif reports RCE.
| Component | Purpose |
|---|---|
| Runtime inventory | Offline exports or a local diagnostic JVM; version, CPU features and effective flags; JSON output and explicit unknown results |
| Semgrep rules | Ten Java rules for decoder inventory, destination size/lifetime review and Spring/protobuf/SnakeYAML entry points |
| Lightweight source search | Optional ripgrep search for Java/Kotlin call sites, including multiline calls |
| ModSecurity indicators | Nine log-only HTTP detectors for selected malformed Base64 shapes; optional JSON parser setup, synthetic tests and explicit coverage limits |
| Experimental IDS module | Suricata/Snort alert-only indicators, offline PCAP tests, exact SID checks and explicit gaps; adapted from jsik22 under MIT |
| Technical analysis | Root cause, CPU gating, allocation behavior, JIT effects and exploitation limits |
| Surface matrix | Product/library paths, version caveats and evidence level |
| Response guide | Updates, verified interim JVM options, operational verification and telemetry |
| Vendor status | Dated advisory and PR checks; distinction between proposals, mitigations and verified shipped fixes |
| Evidence | Sanitized control logs, provenance, supported conclusions and open questions |
Python 3.9+ is sufficient for the inventory tool; it has no third-party dependencies.
# Inspect a trusted Java binary in a NEW diagnostic JVM (not a running service).
python3 safechecks/inventory_j4s.py --java /path/to/jdk/bin/java --json
# Offline assessment: all three exports must come from the same target environment.
python3 safechecks/inventory_j4s.py \
--version-file /path/to/exports/version.txt \
--cpuinfo /path/to/exports/cpuinfo.txt \
--flags-file /path/to/exports/flags.txt --json
# Synthetic example: expected exit code 1 (runtime prerequisites present).
python3 safechecks/inventory_j4s.py \
--version-file examples/version.txt --cpuinfo examples/cpuinfo.txt \
--flags-file examples/flags.txt --json
# Review your source. Semgrep is an optional, separately installed dependency.
semgrep scan --config semgrep/java4shell.yaml --metrics=off /path/to/source
# Java/Kotlin text search without Semgrep; requires rg (ripgrep).
bash safechecks/scan-source.sh /path/to/sourceFor WAF integration, see the ModSecurity deployment guide.
The supplied rules default to pass + log; they do not prove backend exposure or
replace the JVM update. JSON inspection requires phase-1 body/parser setup.
An optional experimental WAF file adds six detectors for query/form values,
cookies, later JWT segments, selected custom headers and plain-text bodies.
The separate IDS module complements these checks; it is not enabled by the WAF
include and has different coverage and false-positive tradeoffs.
Inventory exit codes: 0 = modeled path excluded or mitigated, 1 = runtime prerequisites present, 2 = unknown/incomplete. Exit 1 is not proof of network reachability or exploitation. Exit 0 is scoped to this defect and the supplied runtime evidence; it is not an assertion that the application is secure.
- The accelerated decoder can write a full 64-byte vector at a short padded tail. Part of that store exceeds the destination allocation; the short forms studied produced up to 62 bytes beyond a normally allocated destination.
- A caller-supplied, long-lived destination made adjacent live-object corruption repeatable in the laboratory. This is an important custom-code review pattern.
- Common product call sites use a newly allocated destination. Corruption can be externally silent because the store may land in unused allocation space. No visible crash does not mean no corruption, and does not prove safety.
- Product-level paths were exercised in Elasticsearch, Tomcat and Spring Security laboratory configurations without an observed crash in those runs. Static call-site evidence for other products is labeled separately.
- The fixed JDK 21 control and the disabled-intrinsic control preserved the adjacent canary; the vulnerable default control corrupted it, three runs each.
- Both
-XX:-UseBASE64Intrinsicsand-XX:UseAVX=2prevented the observed store in the campaign's mitigation checks.
The following is the campaign's version model, not a replacement for your JDK supplier's advisory or backport records:
| HotSpot line | Modeled affected releases | Fix baseline used by this pack |
|---|---|---|
| 17 | 17.0.6 through releases before 17.0.21.1 | 17.0.21.1 |
| 21 | 21.0.0 through releases before 21.0.13.1 | 21.0.13.1 |
| 25 | 25.0.0 through releases before 25.0.5.1 | 25.0.5.1 |
| 27 | 27.0.0 through releases before 27.0.1.1 | 27.0.1.1 |
| 18/19/20/22/23/24/26 | Historical affected non-LTS lines | Verify supplier backports or migrate |
| Community 8 / 11 | Investigated upstream trees lack the affected decode intrinsic | Outside this mechanism |
A fourth version component alone does not mean fixed: 25.0.4.1 was a vulnerable laboratory build. Oracle's advisory explicitly lists 8u511b33-perf as affected. This separate 8u-perf lineage is not covered by the community-8 exclusion; the inventory deliberately returns unknown for that label and requires manual vendor-build verification. Unknown does not mean unaffected. Custom builds and backports require package/vendor evidence beyond a version string.
Hardware requirements are real: AVX-512 alone is insufficient; the vulnerable branch requires VBMI and BW visible to the JVM. ARM/AArch64 and AVX2-only environments do not enter this branch. VM/cloud CPU masking matters. OpenJ9 and GraalVM Native Image do not use this HotSpot intrinsic; a GraalVM JVM distribution must not be confused with a native image.
Only the JDK Basic/URL decoding paths are in scope. getMimeDecoder() and the
streaming wrap(InputStream) route do not use the affected path in the reviewed
implementation. A product using another Base64 codec is not automatically affected
through that codec.
- Inventory the actual service JVM, including bundled JDKs and container images.
- Update to your supplier's fixed build and restart the affected processes; prioritize exposed services and the applications in Calif's RCE report.
- If an update is pending, evaluate
-XX:-UseBASE64Intrinsics; check the effective setting in the service after restart. Measure performance on your workload. - Review untrusted-input paths and especially
decode(src, dst)with reused destination arrays. A source-rule finding needs manual data-flow review. - Use operational telemetry for investigation, not as a vulnerability oracle. This pack does not ship a remote trigger, crash scanner or a claim of a reliable exploit-attribution signature. Response details.
python3 -m unittest discover -s tests -v
python3.12 -m venv .venv
.venv/bin/python -m pip install -r requirements-dev.txt
SEMGREP_SEND_METRICS=off SEMGREP_ENABLE_VERSION_CHECK=0 \
.venv/bin/semgrep --test --config semgrep tests/semgrepThe tests exercise classification boundaries, missing evidence, mitigation flags, CLI failure semantics, source-rule positives/negatives and the shipped WAF patterns. The isolated ModSecurity test uses a static Apache server and no Java backend. These checks do not execute the vulnerability. Validation details.
- Calif: Java4Shell — Write once, RCE everywhere
- OpenJDK vulnerability advisory, 2026-11-17
- Oracle Critical Security Patch Update, November 2026
- OpenJDK JDK-8393212 backport: Base64 optimization improvements
- JDK 17u Base64 implementation
- JDK 17u x86 stub generator
- OpenJDK vulnerability advisory index
- Companion PHP defense-tool repository
- jsik22/java4shell-detection: source of the experimental IDS rules and PCAP fixtures
(c) 2026 АО «Сайбер ОК» / JSC CyberOK. See LICENSE for the copyright and usage notice. This repository follows the companion project's notice. The separately licensed experimental IDS module is MIT-licensed and retains upstream attribution.