The security firm behind last month’s high-profile PHP2Shell discovery has used AI to uncover another serious vulnerability, this time in Java. Calif says its researchers have demonstrated remote code execution in widely used Java applications, including Jenkins and JFrog Artifactory.

The vulnerability, called Java4Shell and tracked as CVE-2026-114938, involves an out-of-bounds memory write in OpenJDK’s hardware-accelerated Base64 decoding implementation. Calif disclosed the flaw Tuesday, the same day Oracle released security patches.

Calif reported achieving remote code execution without authentication in default configurations of Jenkins, Artifactory, Keycloak, Jira, Confluence and PingFederate. The firm also demonstrated attacks requiring authentication against several other applications, including Gerrit, Apache Kafka and Elasticsearch.

According to Calif’s technical analysis, the vulnerability affects a Java runtime optimization that requires x86-64 processors supporting AVX-512 VBMI instructions. Independent research from cloud security firm Upwind also found that exposure depends on whether an application provides a path to the vulnerable decoder. That makes the underlying JDK version, processor capabilities and application behavior critical to assessing risk.

How Java4Shell Works

Applications using the standard java.util.Base64.Decoder API can reach the vulnerable routine on affected JVMs.

“The flaw is in OpenJDK’s hand-written assembly for Base64 decoding, which is meant to speed up certain CPUs,” said Ramy Rahman, senior principal solutions engineer at ArmorCode. “Java’s memory safety protections don’t prevent memory corruption in that low-level runtime code.”

HotSpot’s just-in-time (JIT) compiler can select the optimized decoder on compatible processors. Calif found that malformed Base64 padding causes an incorrect write mask, leading the decoder to write beyond its allocated buffer and corrupt neighboring heap memory. Java eventually rejects the malformed input, but the memory corruption has already occurred.

The flaw can also be exposed through authentication processing. As Red Hat explained in a security bulletin for Java4Shell, authentication systems may decode externally supplied credentials or tokens before verifying a user’s identity, potentially allowing unauthenticated attackers to trigger the flaw.

To turn the memory corruption into remote code execution, Calif said its team used heap grooming to manipulate how Java objects are arranged in memory, then developed techniques to read heap contents and gain greater control over memory writes. The firm noted that there is no universal exploit and that successful attacks require techniques tailored to individual applications.

Calif researcher Quang Luong used OpenAI’s Codex to search for vulnerabilities in the JVM’s compiler, optimized routines and cryptographic code. After several unsuccessful leads, Codex identified the memory corruption flaw in the Base64 decoder. Calif said Luong discovered the vulnerability in approximately an hour and developed the first working exploit within another two hours.

The New York Times reported that the research involved GPT-6 Cyber, the specialized OpenAI cybersecurity model available through Codex, and said it independently verified the vulnerability using proof-of-concept details supplied by Calif.

Patching Vulnerable Java Environments

According to Calif, the vulnerability affects unpatched OpenJDK builds starting with Java 17.0.6, including affected releases of Java 21, 25 and 27. Oracle assigned the vulnerability a CVSS 3.1 severity score of 8.1 and classified it as remotely exploitable without authentication, although with high attack complexity.

Organizations running affected Java distributions should obtain patched builds from their vendors. As a temporary mitigation, Red Hat recommends starting affected JVMs with -XX:-UseBASE64Intrinsics to disable the vulnerable optimization.

Red Hat also advises organizations to inventory deployed Java runtimes, including those bundled with applications and containers. Organizations should rebuild affected containers after updating their base images and follow vendor guidance for products with bundled JVMs. Updating application dependencies alone will not resolve the underlying vulnerability.