Skip to content
Neural Daily

ResearchNews

Java4Shell: AI-found critical flaw in OpenJDK

🇭🇺 Olvasd el magyarul →

Security firm Calif disclosed Java4Shell (CVE-2026-114938) on November 17, a memory-corruption bug in OpenJDK's Base64 decoder. It enables unauthenticated remote code execution on Jenkins, Jira, Confluence and Keycloak, among others. The flaw was found with AI help in about an hour.

· 3 min read

Szerver előlapja közelről, perforált fémburkolattal és kék LED-jelzőfénnyel egy adatközpontban
Image: Pexels / Cookiecutter · Pexels · Source

What is the bug?

On x86-64 CPUs, OpenJDK speeds up Base64 decoding with AVX-512 VBMI instructions. According to Calif, the bug sits in that hand-written assembly (Calif). Given 64 A characters followed by a single =, the routine computes a negative output length. That turns into a full 64-byte write, 62 bytes past the end of a 50-byte array. Java rejects the input in the end, but by then neighboring objects may already be corrupted.

With careful exploitation this becomes a Heartbleed-style memory leak or remote code execution. It scores 9.2 under CVSS 4.0; Oracle rated it 8.1 under CVSS 3.1.

What is affected?

Calif demonstrated attacks on default installations. It got pre-auth RCE on Keycloak, Jira, Confluence, JFrog Artifactory, Jenkins and PingFederate, and post-auth RCE on Kafka brokers, Solr, Elasticsearch, Gerrit and Spinnaker. There is a client-side exploit too: a Java program opening a TLS connection to a malicious server is enough.

Red Hat notes that exposure depends on whether an application decodes attacker-controlled Base64 (Red Hat). That is common: HTTP Basic credentials, SAML artifacts, CSRF tokens and MIME encoded words are all Base64, and are often decoded before authentication.

  • Affected: OpenJDK 17.0.6 and newer, plus distributions built on it (Corretto, Temurin, Microsoft Build of OpenJDK, Zulu, Liberica, SapMachine, Oracle Java SE, GraalVM).
  • Hardware condition: x86-64 CPU with AVX-512 VBMI (Intel Ice Lake and newer Xeons, AMD Zen 4 and newer).
  • ARM64 machines are not affected.

What to do now

OpenJDK and Oracle shipped fixes on November 17 (OpenJDK); fixed builds are 17.0.21.1+4, 21.0.13.1+4, 25.0.5.1+4 and 27.0.1.1+4. If you cannot update right away, restart the JVM on x86-64 hosts with -XX:-UseBASE64Intrinsics to disable the accelerated Base64 path, or with -XX:UseAVX=2. Do not set these flags on ARM, where the JVM will refuse to start.

On Linux, the flags in /proc/cpuinfo tell you quickly whether a host qualifies: it needs avx512f, avx512bw, avx512vbmi and bmi2. For containers, rebuild your images once the base image is updated.

Where AI comes in

Calif researcher Quang Luong found the bug in an hour and wrote the first working exploit in two more hours. The company says total token cost stayed under $200. BleepingComputer reports that OpenAI's Codex was used (BleepingComputer), while the New York Times story names the GPT-6 Cyber model (Seattle Times). Calif says later product-specific exploit chains were built almost autonomously by agents.

This is Calif's second major AI-assisted find in a few weeks, after the PHP flaws (OpenAI's Patch the Planet: AI finds critical PHP flaws). It fits the trend in Google GTIG's latest report (GTIG: an AI framework wrote exploits with no human): finding and weaponizing bugs now takes hours, while patch windows have not shrunk.

Key takeaways

  • CVE-2026-114938: memory corruption in OpenJDK's AVX-512 Base64 decoder.
  • Pre-auth RCE on default Jenkins, Jira, Confluence and Keycloak installs.
  • Fixes available since November 17; interim mitigation: -XX:-UseBASE64Intrinsics.
  • Found with AI in about an hour, for under $200 in tokens.

Related articles