A heap overflow in OpenJDK's hardware-accelerated Base64 decoder gives attackers a route to unauthenticated code execution in Jira, Confluence, Jenkins and Keycloak on modern x86 servers. The fix shipped on November 17. Knowing where to apply it is the slow part.
Key Takeaways
Every Java team knows the drill after Log4Shell: find the vulnerable library, bump the version, redeploy. Java4Shell does not live in a library your build file names. It lives in the Java runtime itself, in a decoder that almost every enterprise Java application calls without anyone choosing to, and it only misbehaves on certain processors. That combination makes the patch the easy part. The hard part is producing an honest list of every JVM your organisation runs, including the ones that arrived inside someone else's product.
The flaw, tracked as CVE-2026-114938 and named Java4Shell by the researchers at Calif, is a memory corruption bug in the HotSpot intrinsic that speeds up java.util.Base64 decoding using AVX-512 VBMI instructions. Calif's write-up shows that 64 letter A characters followed by a single equals sign, 65 bytes in total, makes the decoder allocate a 50-byte array and then write 64 bytes into it, overflowing 62 bytes past the end. On OpenJDK 17 the trigger is 128 A characters and an equals sign. Calif rates it 9.2 under CVSS 4.0, and Oracle and OpenJDK disclosed and fixed it on November 17.
The processor dependency is what makes this unusual. VulnCheck found the vulnerable code path runs only on CPUs with AVX-512 VBMI and BW support, and every other CPU decodes through a different path without the defect. In practice that means current Xeon server generations such as Ice Lake and Sapphire Rapids, and AMD Zen 4 and Zen 5 parts including EPYC Genoa and Turin. Most modern Intel desktops and laptops are not exposed, because AVX-512 has been fused off since the 12th-generation Alder Lake chips, and ARM-based Graviton and Ampere servers are unaffected. The same container image can be safe on one host and exploitable on the next, depending on which instance family the scheduler picked that day.
Base64 is everywhere in the request path. Help Net Security relays SentinelOne's point that any value handed to the standard decoder can carry the payload, including HTTP Basic credentials, SAML artifacts, CSRF tokens and MIME encoded words. Calif reports pre-authentication code execution in default configurations of Keycloak, Atlassian Jira and Confluence, JFrog Artifactory, Jenkins and PingFederate, and post-authentication code execution in Kafka brokers, Solr, Elasticsearch, Gerrit and Spinnaker. A client-side variant shared by samczsun compromises a Java program that simply opens a TLS connection to a malicious server.
The severity debate is worth reading carefully rather than averaging. Oracle scores the flaw 8.1 under CVSS 3.1, citing high attack complexity. VulnCheck calls denial of service trivial wherever an attacker can reach a vulnerable decode call: a single request usually did not crash a default server in its testing, but a burst of requests reliably did. Remote code execution is a much higher bar, needing heap grooming and a separate memory leak to defeat ASLR, and VulnCheck expects narrow but working paths in specific applications rather than a universal exploit. Calif agrees there is no single exploit, and says each target needs its own application profile.
Nobody had reported exploitation in the wild when VulnCheck published, and Kevin Beaumont told readers the flaw is not world ending and can follow normal patch cycles. That reassurance applies to opportunistic mass scanning. It does not cover the denial-of-service risk to an internet-facing identity provider, and it sits awkwardly beside the timeline Calif describes. Its researchers say an AI model, OpenAI's Codex, produced the Base64 finding after about 20 minutes of persistent prompting, discovery took roughly an hour, and the first working exploit followed two hours later, at a total token cost under $200. Calif is withholding its product-specific chains, but VulnCheck says its own full exploit went to paying customers at 6 PM Eastern on the day of disclosure.
“AI brings the horsepower. Our researchers bring the context.”
Calif
The fix exists. Oracle and the downstream builders, including Amazon Corretto, Eclipse Temurin, Microsoft's OpenJDK build, Azul Zulu, BellSoft Liberica, SAP SapMachine and Red Hat, are shipping fixed builds. The question is where to install them. The CISA alert issued on November 17 tells organisations to inventory every JDK and Java runtime, including those bundled with third-party applications, container images and appliances, and warns that these may not appear as separate inventory entries at all. A Jira or Artifactory install that carries its own runtime will not be fixed by updating the host's Java package. A base image pinned last spring will keep rebuilding the vulnerable JVM until someone changes the tag. A vendor appliance will wait on the vendor's release schedule.
That gap is the real exposure window. Software composition tooling built around build manifests tends to see the libraries an application declares, not the runtime it executes on, and an appliance gives security teams no manifest to read. The hardware dimension makes prioritisation harder still, because the honest answer to whether a given service is exploitable depends on the CPU flags of the host it happens to run on. On Linux, Calif says the decisive flags in /proc/cpuinfo are avx512f, avx512bw, avx512vbmi and bmi2.
Log4Shell taught the industry to keep a bill of materials for its libraries. Java4Shell tests whether that discipline reaches one layer lower, to the runtimes that vendors embed and base images freeze in place. With AI cutting the time from bug to working exploit to an afternoon, the inventory that took weeks to build in 2021 now needs to exist before the advisory lands, not after it.
Guide
Java4Shell is a malformed-input bug in an optimised decoder that only fires on one CPU family, exactly the kind of edge case this guide argues fuzzing finds when conventional scanners and code review walk straight past.
Download
Guide
CISA warns that bundled Java runtimes may never show up as separate inventory entries, and this buyer's guide lays out what an enterprise tool needs to see the open source and third-party components your applications quietly ship with.
Download
Whitepaper
Appliances and packaged products that embed their own JVM can only be fixed when the vendor ships a new build, and this whitepaper shows what a vendor's incident response and patch process should look like when a flaw lands in a component it bundles.
Download
Two rogue agent-framework releases on PyPI ran a credential stealer at every Python start, even where the framework was never imported.
Sonatype found a loader in 12 hijacked npm releases, and none of them used an install hook a script policy could block.
A master key leaked in May and never revoked let attackers run LiteLLM's own release pipeline in October.