"Successful attacks of this vulnerability can result in takeover of Oracle Java SE," Oracle wrote in its CVE record — a stark legal-sized sentence that frames the risk disclosed this week as CVE-2026-114938, nicknamed Java4Shell by the researchers who found it.
Calif's disclosure and the scope of affected software
Security firm Calif disclosed the vulnerability and published technical details showing that unauthenticated attackers can run their own code against multiple widely used server applications when certain Java builds run on particular x86-64 processors. Calif rates the flaw 9.2 out of 10 under CVSS 4.0. Oracle, Red Hat and Amazon have issued differing CVSS scores: Oracle 8.1 (CVSS 3.1), Red Hat a preliminary 9.8, and Amazon 8.8.
Calif says the vulnerability allows pre-auth remote code execution by default in Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate. With a login required, Calif reports exploitation is possible in Apache Kafka Broker, Apache Solr, Elasticsearch, Gerrit and Spinnaker. The researchers are withholding full exploit code for the six unauthenticated targets; other labs have reproduced crashes and a memory leak but not confirmed remote code execution.
AVX-512 VBMI Base64 decoder: where the bug lives
The bug is in a faster version of Java's Base64 decoder that Java uses on chips supporting AVX-512 VBMI. It appears in Java 17.0.6 and later on the Java 17 line, and in Java 21, 25 and 27. Standard Java 8 and 11 do not contain it, though the Java SE 8 Performance Pack build 8u511b33-perf is affected. GraalVM builds for certain JDK versions are also listed as vulnerable.
The vulnerable code runs only on x86-64 processors that support avx512f, avx512bw, avx512vbmi and bmi2 — for example, Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD Zen 4 or newer. ARM64 chips do not run the vulnerable path. Calif notes that the same container image can be vulnerable on one host and not on another, depending on the CPU features the server presents.
On Linux, the source provides a command to check whether all four CPU flags are present: grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' | grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'.

This site is the portfolio.
OSINTSights runs on Cloudflare Workers, D1, R2, and Vectorize, with an AI pipeline on Hetzner ARM. Nubivance designed, built, and operates it. We do the same for clients.
See what we buildHow the exploit works: the overflow and JavaBleed
Calif describes a malformed Base64 string that ends with an extra '=' which causes the AVX-512 decoder to write 62 or 63 bytes past the end of its output buffer. Java rejects the string, but only after the out-of-bounds write occurs. The bytes written past the end are fixed; early readings of the source code suggested a crash rather than code execution because an attacker controls only the placement, not the content, of those bytes.
To get past that limit, the researchers arrange server memory so the overflow expands a neighboring array. They then use that array to write arbitrary bytes and a second array to read back memory — a step they call JavaBleed. From that point, Calif says, the attack can reach parts of the JVM that run code. There is no single, universal exploit: each product requires a bespoke memory layout and precise timing, and some variants are sensitive to network delays. A proxy that waits for complete requests, such as nginx in its default configuration, blocks some attack attempts but not all.
Reproductions and lab results: Upwind, Blackpoint Cyber, and others
Independent labs have reproduced memory corruption and crashes. Upwind reproduced the memory corruption on an AWS instance with an Intel Ice Lake processor, observing damage to the next object in memory and eventual JVM crash. Blackpoint Cyber's Adversary Pursuit Group crashed a test Spring Boot application by sending roughly 28,300 malformed login requests over about 20 seconds, without logging in. In local tests Blackpoint used JavaBleed to read a secret stored adjacent to the overflow; that local read succeeded in all 30 runs but the group did not reproduce remote code execution.
Canada's Cyber Centre said it was not aware of any attacks exploiting the flaw. The flaw did not appear in the KEV catalog of exploited vulnerabilities maintained by the U.S. Cybersecurity and Infrastructure Security Agency when The Hacker News checked.
Vendor fixes, patches, and interim workarounds
Oracle and the OpenJDK project released fixes this week. Oracle's .1 release updates for 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1 include the security fixes; OpenJDK lists the issue as fixed in this week's releases. Amazon's Corretto builds and Amazon Linux packages are marked fixed; Azul updated Zulu builds for 17, 21, 25 and 27; GraalVM published November security releases. Red Hat has not yet released fixes for RHEL 8, 9 and 10 or for its OpenJDK 17, 21 and 25 builds; Red Hat warns container images based on its images must be rebuilt when fixed images appear. SAP released SapMachine builds this week, though its release pages do not specify the fixes. Eclipse Temurin, the Microsoft Build of OpenJDK and BellSoft Liberica had not listed fixed builds when The Hacker News checked.
If immediate updating is not possible, the OpenJDK advisory lists two JVM startup flags as mitigations: -XX:-UseBASE64Intrinsics disables only the faster Base64 code path, removing the vulnerable path while leaving other AVX-512 speedups; -XX:UseAVX=2 turns off AVX-512 entirely (use only on x86-64 servers). Upwind found both flags stopped the overflow in its tests. A different flag, -XX:DisableIntrinsic=_base64_decodeBlock, can prevent Java from starting if used alone.
Upwind recommends looking for base64 tokens of 32 characters or more that end in '=' whose length is not a multiple of four (especially in Authorization headers), bursts of IllegalArgumentException errors from java.util.Base64$Decoder, and unexplained hs_err_pid crash logs or service restarts — while cautioning none of these signs proves compromise on their own.
What this means for technologists, product vendors, and operators
- Technologists and security teams should inventory Java versions and CPU features: vulnerable Java builds and AVX-512 VBMI-enabled hosts are a necessary conjunction for this exploit to run.
- Product vendors that bundle their own JRE must issue updates; Calif and vendor advisories stress that products bundling Java runtimes need vendor-supplied updates and container images rebuilt and redeployed.
- Operators of internet-facing services listed by Calif should prioritize testing patched builds or applying the suggested JVM startup flags until vendor fixes are available — noting that some mitigations may impact performance or startup behavior.
The technical ingenuity of JavaBleed and the narrow hardware precondition make Java4Shell a complex but serious fault: patches are available and labs have reproduced damaging crashes and memory reads, yet public exploit code remains withheld and independent teams have not broadly reproduced remote code execution. For now, the clear actions are inventory, patch, or apply JVM mitigations — and watch whether proof-of-concept code appears outside the researchers' controlled disclosure.




