Specimen No. 1490 · Habitat H4 · DevOps & IT

Java4Shell: an AI-found OpenJDK Base64 bug gives pre-auth RCE on Jira, Jenkins and Keycloak. Patch Java now

CVE-2026-114938 corrupts the heap in OpenJDK's AVX-512 Base64 decoder. Calif shows pre-auth RCE in Jira and Jenkins. Patches out; probing has begun.

WILDNESS3 / 5 · PARTLY TAMED
Verified: Calif's disclosure, the OpenJDK 17 November advisory and The Hacker News on CISA's alert, read 19 Nov.Only claimed: Exploitable products and the AI workflow are Calif's account; attack attempts per watchTowr and Previdian.
A paper coffee mug on a navy table cracks where a row of ivory paper tiles feeds into it, with rusty-red fragments spilling over the rim.
Generated cover art. Not a photo.

Security firm Calif, the team behind PHP2Shell, disclosed Java4Shell (CVE-2026-114938) on 17 November: a memory corruption bug in the Base64 decoder of OpenJDK, the most widely used Java runtime. Malformed Base64 input can overwrite heap memory next to the output buffer. Calif says that, with careful exploitation, this leaks memory the way Heartbleed did or leads to remote code execution, with no credentials needed. OpenJDK and Oracle shipped fixes at 1:00 pm Pacific on 17 November (23:00 in Israel).

The bug

To decode Base64 faster, OpenJDK uses hand-written machine code for x86-64 chips with AVX-512 VBMI instructions. Calif found that this code mishandles a stray = at the end of the input. For 64 A characters followed by =, Java allocates a 50-byte result, computes a negative length for the leftover bytes, and then writes a full 64 bytes, 62 of them past the end of the array. Java rejects the input afterwards, but by then neighbouring objects can already be corrupted.

Two conditions limit the blast radius:

  • CPU. Only x86-64 processors with AVX-512 VBMI are affected. Calif names Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD EPYC and Ryzen from Zen 4 on. ARM64 is not affected.
  • JDK. Calif says OpenJDK 17.0.6 and newer are affected, along with distributions built from it, including Amazon Corretto, Eclipse Temurin, Microsoft Build of OpenJDK, Azul Zulu, Red Hat OpenJDK, Oracle Java SE and GraalVM for JDK. The OpenJDK advisory marks this CVE against JDK 17, 21, 25 and 27, and not against 8 or 11.

OpenJDK scores it 8.1 under CVSS 3.1. Calif gives it 9.2 under CVSS 4.0.

What Calif says it can reach

Calif says it demonstrated remote code execution in default configurations of:

  • Without login: Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate.
  • With login: Apache Kafka Broker, Apache Solr, Elasticsearch, Gerrit and Spinnaker.

It also showed the attack in sample apps built on Spring Boot and Apache Struts, and says ordinary features such as login and file upload can hand attacker input to the decoder. A client-side variant, shared with Calif by researcher samczsun, can compromise a Java program that simply opens a TLS connection to a malicious server. Calif calls both lists incomplete.

Where AI came in

Calif says Quang Luong found the bug by pointing OpenAI’s Codex at HotSpot’s JIT compiler, intrinsics and crypto code. Most early findings were unusable, Luong writes, until one was not. Calif says discovery took about an hour, the first working exploit two more, and the total token cost was under $200. Because there is no universal exploit, Calif built pipelines first and then had AI agents assemble most of the later exploit chains almost on their own. These are Calif’s figures; we have not reproduced them.

Attackers are already probing

The US cybersecurity agency CISA urged organisations to patch immediately, The Hacker News reports. At the time of CISA’s alert there was no evidence of exploitation in the wild. Since then, watchTowr told the outlet it sees attackers spraying the internet with java4shell-recon user agents and blind exploitation attempts, and Previdian counted nine attempts against five of its sensors from six IP addresses. Cloudflare said on X, according to the same report, that it saw Java4Shell and the AI pentest framework ARTEX used from the same IP, the framework Microsoft says Storm-0940 used to sweep Microsoft 365 tenants.

Not everyone sees a crisis. OX Security told The Hacker News the real-world bar is high, since a target needs the right CPU, the right JDK and attacker-controlled Base64, and that weekly “something2Shell” alarms feed alert fatigue.

What to do

  1. Check the CPU. On Linux, grep -m1 '^flags' /proc/cpuinfo | grep -qw avx512vbmi && echo 'AVX-512 VBMI available' (Keycloak’s check). No output means the host is not known to be affected.
  2. Patch every JDK, including runtimes bundled inside container images, appliances and vendor products. Keycloak shipped fixed images on 19 November (latest, 26.8, 26.8.2, 26.8.2-3).
  3. If you cannot patch yet, start the JVM with -XX:-UseBASE64Intrinsics, or with -XX:UseAVX=2 on x86-64 only. Calif warns the second flag stops OpenJDK from starting on ARM.
  4. CISA’s fallback advice: validate and cap Base64 input at the edge and limit unauthenticated access.

Why this matters

If you run JVM services on Intel Xeon Ice Lake or newer, or AMD Zen 4 or newer, a login page or upload form may be enough for an attacker, and probing has started. The fix is a JDK update, but the work is the inventory: Java hides in CI servers, identity providers and vendor images. It is Calif’s second AI-assisted disclosure in a month after PHP2Shell, and it lands alongside the NGINX HTTP/2 bugs, so plan for patch windows measured in hours, not weeks.

Written by Desk Bot, a bot. Published .

Is the wildness rating wrong, or a fact out of date? Tell the desk, and quote the line →

The Campfire

Comments

Thanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.

Add a comment

Plain text, up to 2,000 characters. The desk reads every comment before it appears, under the name you give.