TL;DR
- what: Calif disclosed Java4Shell (CVE-2026-114938), an out-of-bounds write in Java's AVX-512 VBMI Base64 decoder that it turned into remote code execution with no login.
- impact: Default installs of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate are exploitable pre-auth, five more products with a login, and anyone can crash vulnerable JVMs with malformed Base64.
- fix: Oracle and OpenJDK fixed it on November 18 (Oracle builds 17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1 and Java 8 Performance Pack 8u513), Amazon Corretto is patched, and Red Hat builds are still pending.
- who: Any service on Java 17.0.6 or later, 21, 25 or 27 running on Intel Ice Lake or newer Xeon, or AMD Zen 4 or newer, that decodes Base64 from untrusted input.
Security research firm Calif disclosed Java4Shell (CVE-2026-114938) this week. It is a memory corruption flaw in Java's hardware-accelerated Base64 decoder, and Calif says it turned it into unauthenticated remote code execution on the default configurations of Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate. Oracle and the OpenJDK project released fixes on November 18. Calif scores the flaw 9.2 under CVSS 4.0 and says its list of affected applications is incomplete.
There are real limits on exploitation. The vulnerable code only runs on x86-64 CPUs that support AVX-512 VBMI. Calif is withholding its exploits for the six applications. Other labs have reproduced crashes and a memory leak, but not code execution. Canada's Cyber Centre says it is not aware of any exploitation, and the flaw was not in CISA's Known Exploited Vulnerabilities catalog as of November 18.
How the Bug Works
Servers routinely decode Base64 before they check whether a user is logged in. HTTP Basic authentication headers, SAML artifacts and CSRF tokens all arrive Base64-encoded and get decoded first. That ordering is what makes the flaw reachable without an account.
The defect is in a faster version of the decoder that Java switches to on AVX-512 VBMI hardware once the JIT compiler has optimized the decoder, which it does for code that runs often. On a busy server that happens quickly. A malformed string ending in an extra = character makes this decoder write 62 or 63 bytes past the end of its output buffer. Java does reject the string, but only after the write has already happened.
The overflow bytes are fixed, so an attacker controls only where they land, not what they contain. That is why early readings of the source code, including The Hacker News' first report, pointed to a crash rather than code execution. Calif got past this by arranging heap memory so the overflow enlarges a neighboring array. It uses that array to write bytes of its choosing and a second array to read memory back, a step it calls JavaBleed. From there, Calif says, the attack reaches the parts of the JVM that execute code.
⚠️ Fixed bytes did not mean crash only — The overflow content is fixed, but Calif turned the overflow's location into arbitrary read and write on the heap. In practice, any reachable Base64 decode of attacker input on vulnerable hardware should be treated as a possible code execution path, not just a denial of service.
Exploitation Is Real but Target-Specific
- There is no universal exploit. Each application needs its own memory layout and timing, and some attacks are sensitive to network latency.
- A reverse proxy that buffers complete requests, as nginx does by default, blocks some attacks but not all of them.
- Failed attempts can crash the JVM, and attackers can often just try again.
- With a login, Calif also reached code execution on Apache Kafka Broker, Apache Solr, Elasticsearch, Gerrit and Spinnaker.
- Calif says a Java client that opens a TLS connection to a malicious server can be compromised even if it runs no service of its own. That exploit came from researcher samczsun, and the path to the decoder has not been published.
What Independent Labs Confirmed
Upwind reproduced the memory corruption on an AWS instance with an Intel Ice Lake CPU. The overflow damaged the next object in memory and eventually crashed the JVM. Blackpoint Cyber's Adversary Pursuit Group crashed a test Spring Boot application with about 28,300 malformed login requests over roughly 20 seconds, without ever logging in. In a local test, the group also used the JavaBleed step to read a secret stored next to the overflow, and it worked in all 30 runs. Neither lab reproduced remote code execution.
Vendors Disagree on Severity
Oracle scores the flaw 8.1 under CVSS 3.1, rating it hard to exploit but requiring no login, and its CVE record warns that successful attacks "can result in takeover of Oracle Java SE." Red Hat's preliminary score is 9.8, also with no login required. Amazon scores it 8.8 and describes the risk as "a JVM crash for unauthenticated users or, in certain application environments, code execution by an authenticated low-privileged user." Canada's Cyber Centre calls it remotely exploitable without authentication.
The spread comes down to one question. Every vendor accepts that an unauthenticated attacker can crash the JVM. Where they differ is how reliably that becomes unauthenticated code execution in real deployments.
Who Is Exposed
- Java 17.0.6 and later on the 17 line, plus Java 21, 25 and 27.
- Standard Java 8 and Java 11 do not contain the vulnerable code, but the Java SE 8 Performance Pack build 8u511b33-perf does.
- GraalVM for JDK 17.0.21.0.3, GraalVM for JDK 21.0.13.0.3 and GraalVM Enterprise Edition 21.3.20.0.3.
- Only x86-64 CPUs that report all four of avx512f, avx512bw, avx512vbmi and bmi2: Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD Zen 4 or newer.
- ARM64 systems do not run the vulnerable code.
Same image, different risk — Because the vulnerable path depends on the CPU, one container image can be safe on one host and exploitable on the next. Mixed cloud instance families, autoscaling groups and Kubernetes node pools can move a workload onto vulnerable hardware with no change to the image or the JDK version.
Vendor Fixes and Mitigations
- OpenJDK: the project's advisory lists the flaw as fixed in this week's releases.
- Oracle Java: the fixed builds are the new .1 releases, 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1, all released this week.
- Java 8 Performance Pack: build 8u513, available in standard and Performance Pack versions.
- GraalVM: fixed in Oracle's November security releases, and Oracle's advisory lists the affected builds.
- Amazon Corretto: fixed. Red Hat: fixes are still in preparation.
- Applications that ship their own bundled Java runtime will need an update from that application's vendor, separate from the host JDK.
- Partial mitigation: request-buffering proxies such as nginx with default settings block some, but not all, of Calif's attack variants.
RedEye Assessment
Treat Java4Shell as a confirmed unauthenticated denial of service today, and as unauthenticated RCE the moment Calif's exploits or an independent reproduction go public. The six pre-auth targets are identity providers, CI servers and artifact stores, systems that hold credentials, signing keys and build pipelines, which puts them at the top of any attacker's list. The CPU dependency makes exposure harder to inventory than a normal JDK flaw, because the version number alone does not tell you whether a host is at risk. Expect researchers to close the gap between crash and code execution quickly now that the JavaBleed technique has been described.
Originally published on RedEye Threat Intelligence.
Top comments (0)