Skip to content
AI.info

The Pulse

AI-assisted research finds critical OpenJDK code-execution flaw

Calif says it used AI to find a memory-corruption flaw in OpenJDK and demonstrate attacks against several Java applications. Oracle has released patched Java builds, while researchers say exploitation depends on the hardware and application.

AI-assisted research finds critical OpenJDK code-execution flaw

AI-assisted research has uncovered a critical OpenJDK flaw that could let attackers execute code remotely or read sensitive memory in vulnerable Java applications, security firm Calif says. The researchers say they demonstrated attacks against several widely used products, including some that required no authentication.

Calif disclosed the flaw, named Java4Shell and tracked as CVE-2026-114938, on November 17. Oracle released patched Java builds the same day.

A decoder flaw

The vulnerability sits in OpenJDK’s Base64 decoder. Base64 represents binary data as text and is used in functions including authentication, file uploads and network messages.

The affected optimization uses AVX-512 VBMI instructions on supported x86-64 processors. Calif says malformed input can make the decoder write beyond an allocated memory buffer, corrupting neighboring objects before Java rejects the input.

In one example described by the researchers, a 65-character input leads OpenJDK to allocate a 50-byte result array. The decoder then miscalculates the output length and performs a 64-byte write, sending 62 bytes beyond the array.

The flaw carries a critical score of 9.2 out of 10 under CVSS 4.0, according to Calif. Oracle assigned it an 8.1, or high severity, under CVSS 3.1, security company Echo says.

Demonstrated attacks

Calif demonstrated unauthenticated remote code execution in default configurations of Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate. It also demonstrated code execution after authentication in Apache Kafka Broker, Apache Solr, Elasticsearch, Gerrit and Spinnaker.

The lists are not exhaustive, and the researchers also demonstrated code execution in sample applications built on Apache Struts and Spring Boot. A separate client-side exploit shared by researcher samczsun showed that a Java program connecting to a malicious server over TLS could be compromised even if it did not expose a service, Calif says.

The company says there is no universal exploit: the steps depend on how an application works and uses Base64. Calif is withholding product-specific exploit details to give users time to upgrade, CyberScoop reports.

How AI was used

Researcher Quang Luong directed Codex to examine parts of OpenJDK including the just-in-time compiler, commonly accessible APIs and cryptography. Calif says Luong found the vulnerability in an hour and developed an initial working exploit in another two hours. The total AI token cost was less than $200.

Luong wrote that many early results from Codex were unreachable or had unrealistic exploitation requirements. After researchers built pipelines and strategies for exploitation, AI agents helped with initial work and built most later exploit chains almost autonomously, according to Calif.

Versions and fixes

Echo says the flaw affects unpatched OpenJDK builds starting with version 17.0.6, including affected releases across Java 17, 21, 25 and 27, on processors supporting the relevant instructions. Certain Oracle Java 8 Enterprise Performance Pack builds are also affected, while standard OpenJDK 8 and 11 releases are not included in the affected range.

As of November 18, there were no confirmed reports of exploitation in the wild, according to Echo. Oracle’s patched builds include fixes for the affected Java releases.

Sources

  1. Calif
  2. CyberScoop
  3. Echo

Related stories

Explore

More articles