Ai Engineering 7 min read

Java4Shell: an AI-assisted audit found a pre-auth RCE in OpenJDK's Base64 decoder for under $200 in tokens

Oracle and OpenJDK patched CVE-2026-114938 on November 17, a heap overflow in HotSpot's AVX-512 Base64 intrinsic that Calif says it found with OpenAI Codex in about an hour and turned into remote code execution in Jira, Confluence, Keycloak and Jenkins. Here is how the bug works, who is affected and what to do.

Oracle and the OpenJDK Vulnerability Group shipped fixes on Tuesday, November 17, for CVE-2026-114938, a memory corruption bug in the HotSpot JVM that the security firm Calif has named Java4Shell. The flaw sits in the hand-written x86-64 assembly that accelerates java.util.Base64 decoding on CPUs with AVX-512 VBMI, and Calif says it has turned it into remote code execution, without credentials, in default installs of Keycloak, Atlassian Jira and Confluence, JFrog Artifactory, Jenkins and PingFederate.

Two things make it more than another Patch Tuesday entry. Base64 decoding of attacker-supplied input is everywhere in Java servers (auth headers, tokens, SAML, JSON payloads), so the reachable surface is enormous. And Calif says the bug was found and weaponized with AI: researcher Quang Luong pointed OpenAI’s Codex at the JIT compiler and intrinsics, and Calif reports about an hour to discover the bug, two more hours to a first working exploit, and a total token bill under $200. It is the clearest example yet of the economics we saw with Broccoli2Shell in October, now landing in the most widely deployed server runtime there is.

How the bug works

HotSpot’s C2 JIT compiler replaces hot calls to Base64.Decoder with an intrinsic, a hand-tuned assembly routine that uses AVX-512 VBMI instructions to decode 64 bytes at a time. Per Calif’s write-up, the error is in how that routine handles the last few characters of padded input. When the input ends in =, the computed remaining-output length underflows to -1 or -2, and the bzhi instruction used to build a store mask turns that into an all-ones mask. The routine then writes a full 64-byte block past the end of the output buffer.

Calif’s minimal trigger is 64 A characters followed by = (65 bytes) on Java 21, 25 and 27, and 128 A characters plus = on Java 17. In its example, a 50-byte output buffer receives 62 bytes of out-of-bounds writes on the Java heap.

The overwrite is not fully attacker-controlled: Calif describes the written bytes as largely fixed, which is why Oracle rates attack complexity as high. Calif’s chains first use the corruption to leak heap memory (it calls that stage “JavaBleed”), then groom the remote heap so the constant overwrite lands somewhere useful. Calif says there is no universal exploit and each product needs its own chain. Crashing a vulnerable JVM, on the other hand, is easy, so denial of service is the realistic floor for any exposed service.

Who is affected

Every major build of OpenJDK carries the same HotSpot code, so the distribution name does not matter; the version and the CPU do.

  • Versions: Calif, Tenable and Endor Labs agree that HotSpot from 17.0.6 onward is affected, along with the 21, 25 and 27 lines. 17.0.0 to 17.0.5 and Java 11 are not. Standard Java 8 is not affected, though Oracle’s own list includes the Java 8 Enterprise Performance Pack build 8u511b33-perf (Endor’s table disagrees on that one build).
  • Hardware: only x86-64 CPUs with AVX-512 VBMI: Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD Zen 4 or newer (EPYC and Ryzen). That is most modern cloud instances. ARM64, including Graviton and Apple silicon, is not affected. Calif’s page gives a /proc/cpuinfo check for the avx512vbmi flag.
  • Distributions: Calif lists Amazon Corretto, Eclipse Temurin, Microsoft Build of OpenJDK, Azul Zulu, BellSoft Liberica, SAP SapMachine, Red Hat OpenJDK, Oracle OpenJDK, Oracle Java SE and Oracle GraalVM for JDK. Upwind says its lab reproduced the bug on Temurin 17.0.6, 17.0.21, 21.0.13 and 25.0.5.

Oracle’s fixed Java SE builds are 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1 (with 8u513-perf for the Performance Pack), per Tenable and Aikido. Aikido lists matching Corretto builds (17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2, 27.0.1.1.4) and says Red Hat is expediting RHEL 8, 9 and 10 updates. Endor reported on November 17 that the fix (JDK-8393212) was tagged in jdk-27.0.1.1-ga but had not yet appeared in the public jdk17u, jdk21u and jdk25u repositories, so builds that compile from upstream may lag. Check your own vendor’s advisory rather than assuming a version number carries over. EOL lines (18 to 20, 22 to 24, 26) get no fix.

Severity: 9.2 or 8.1, and what is actually known

Calif scores the bug 9.2 under CVSS 4.0. Oracle’s advisory, as reported by Tenable and SecurityWeek, gives 8.1 under CVSS 3.1: network-reachable without authentication, but high complexity. Oracle adds that exploitation requires an application to pass untrusted data to the affected Base64 APIs, which in practice describes most web-facing Java software. Oracle credits Quang Luong of Calif and Albert Yang.

As of November 18, no source reports exploitation in the wild, and the CVE is not in CISA’s Known Exploited Vulnerabilities catalog. But Calif is no longer the only group with a working chain. Kevin Breen of Immersive told Tenable he has an application-specific exploit, and VulnCheck told SecurityWeek it has full RCE while predicting the bug will not be “easily or widely exploitable.” Cisco Talos expects probing to start soon after disclosure. Calif says it is withholding product exploits until users have had time to patch, and has published only a reproducer that confirms the memory overwrite without achieving code execution. The working product chains (pre-auth: Keycloak, Jira, Confluence, Artifactory, Jenkins, PingFederate; post-auth: Kafka, Solr, Elasticsearch, Gerrit, Spinnaker; plus Struts and Spring Boot sample apps) are Calif’s claims and have not been independently reproduced. Calif also says samczsun built a client-side variant in which a Java client connecting over TLS to a malicious server is compromised, which widens the exposure beyond servers.

The AI angle: cheap discovery, cheap weaponization

Calif’s account of the discovery is notably unglamorous. Luong scoped the audit to the JIT compiler, intrinsics, widely used APIs and cryptography, then asked Codex to look at those areas. Most early findings were false leads; the Base64 bug surfaced after roughly 20 minutes of prompting and discarding noise, by the write-up’s own description. The team then built a heap tracer and visualizer so the agent could see heap state while developing the exploit, and Calif says AI agents produced most of the subsequent product-specific chains.

The cost figures, under $200 and about three hours from start to first exploit, are Calif’s and unaudited. But they fit a pattern this autumn: AI agents finding bugs in decades-old editors, in Brotli, and now in hand-written JIT assembly that human reviewers had read since the AVX-512 path was backported in early 2023. Hand-optimized intrinsics are a good target for this kind of search because they are small, performance-critical, rarely touched and hard for humans to reason about at the edge cases. Expect other teams to point agents at the remaining HotSpot intrinsics (CRC32, AES, string, array copy) next.

For defenders the lesson is about timelines. If finding and weaponizing a bug in core Java costs a few hundred dollars, the window between a patch landing and a private exploit existing is now measured in hours, as the Immersive and VulnCheck chains a day after disclosure already show.

What to do now

  1. Inventory x86-64 JVMs on 17.0.6 or newer running on AVX-512 VBMI hardware, including vendored runtimes inside appliances and container base images. Jira, Confluence, Jenkins, Keycloak and Artifactory instances exposed to the internet come first.
  2. Upgrade to your vendor’s fixed build and rebuild container images; a patched host JDK does not fix an image that bundles its own.
  3. If you cannot patch today, start the JVM with -XX:-UseBASE64Intrinsics to turn off the vulnerable routine. Endor suggests the narrower -XX:+UnlockDiagnosticVMOptions -XX:DisableIntrinsic=_base64_decodeBlock (Upwind notes the second flag fails to start the JVM without the first). -XX:UseAVX=2 also works on x86-64 but disables AVX-512 entirely and will stop an ARM64 JVM from starting, so do not push it fleet-wide. Expect some slowdown in Base64-heavy services; none of these replace patching.
  4. Watch for crashes. Unexplained JVM segfaults in Base64 paths after November 17 are worth treating as possible probing.

What to watch next: vendor builds for Temurin, Azul and Microsoft OpenJDK landing in the public update repositories, Calif’s release of its product exploits, any CISA KEV listing, and whether Atlassian, Jenkins and the Keycloak project ship their own advisories or bundled-JDK updates.

Get Insanely Good at AI

Get Insanely Good at AI

The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.

Keep Reading

Ai Agents

How to Secure AI Agents With Google ADK

Learn how to secure your autonomous workflows and prevent unauthorized actions using Google ADK's hardware-backed tool binding and execution logs.

Ai Engineering

Broccoli2Shell: the Brotli Bug Gets a Census, and OpenAI Codex Security Gets the Credit

The Broccoli2Shell site published October 25 maps CVE-2026-111497 across every ecosystem: vendored copies in .NET, DuckDB, and Apache Arrow remain unfixed, non-C decoders are clean, and the original bug report came from OpenAI Codex Security.

Ai Engineering

HTTP2Shell: the Brotli Bug Behind PHP2Shell's Naming Season Reaches Node and Python

CVE-2026-111497, an out-of-bounds write in the Brotli decoder bundled with Node.js, escalated October 23 into 'HTTP2Shell' as Aikido confirmed Python HTTP libraries using Brotli are affected too; fixed Node releases shipped October 21 with no workaround for outbound fetch paths.

Ai Engineering

Microsoft's Ignite model drop: MAI Cyber 2 Flash, Code 2, Realtime 1 and Decision 1.1 aim at price, not the frontier

At Ignite on November 17, Microsoft AI announced seven first-party models, from a 1M-context security model priced at $0.50 per million output tokens to coding models in GitHub Copilot and an OpenAI-compatible speech-to-speech model at $0.01 a minute. Here is what each does, where it runs and what the benchmark fine print says.

Ai Engineering

Chinese AI labs are trimming offensive cyber skills from new models as Beijing drafts export curbs

Bloomberg reported on November 16 that Alibaba, Tencent and Moonshot have shipped recent models with reduced offensive cyber capability after a Ministry of Commerce draft proposed banning exports of AI attack tools. The methods differ, and only Moonshot published a number.