Calif: OpenAI’s Codex Found Java4Shell for Under $200

Security firm Calif disclosed Java4Shell (CVE-2026-114938) on 17 November, “a memory-corruption vulnerability in OpenJDK’s Base64 decoder” that it says lets “a remote attacker with no credentials” reach remote code execution. Calif says its researcher found the bug with Codex, OpenAI’s coding agent, in an hour, and that “the total token cost was less than $200”.

Our number: OpenJDK’s own advisory rates the flaw 8.1 under CVSS 3.1, the highest of the 10 OpenJDK CVEs fixed in its 17 November release. It marks the flaw as affecting 4 of the 6 release lines it lists (17, 21, 25 and 27), by our reading of the risk matrix. Calif scores it 9.2 under CVSS 4.0.

Business Pill · Red Teaming

An evergreen Business Engineer explainer on red teaming: attacking your own systems to find weaknesses first. It teaches the general idea only and does not discuss Java, Calif or this disclosure.

The key insight: Calif says a researcher using Codex found Java4Shell, an OpenJDK remote-code-execution flaw, in an hour for under $200 of tokens; OpenJDK rates it 8.1, the highest of the 10 CVEs in its 17 November advisory, by our reading.

What Calif Disclosed

Calif says the bug sits in hand-written assembly that OpenJDK uses to decode Base64 faster on x86-64 chips, and that “malformed input” can “overwrite neighboring memory”. Its write-up says the corruption “can reveal sensitive memory, as Heartbleed did, or lead to remote code execution”.

Calif lists pre-authentication remote code execution in six widely deployed Java applications in their default configurations: Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate. It lists post-authentication RCE in five more, including Apache Kafka, Apache Solr and Elasticsearch, and calls both lists “non-exhaustive”.

Calif says it is “holding back the application-specific exploits for now” and that “the weaponized chains for individual products stay private until everyone has had a chance to upgrade”. It also notes “there is no universal exploit”.

CVSS 3.1 base scores for the 10 OpenJDK CVEs fixed in OpenJDK's 17 November 2026 vulnerability advisory. Java4
CVSS 3.1 base scores for the 10 OpenJDK CVEs fixed in OpenJDK’s 17 November 2026 vulnerability advisory. Java4Shell (CVE-2026-114938, hotspot/compiler) is rated highest at 8.1 and marked as affecting OpenJDK 17, 21, 25 and 27. Calif scores it 9.2 under CVSS 4.0.

Who Is Affected

Calif says Java4Shell “affects OpenJDK from version 17.0.6 or newer” and names ten distributions built on it, including Amazon Corretto, Eclipse Temurin, the Microsoft Build of OpenJDK, Azul Zulu, Red Hat OpenJDK and Oracle Java SE. It adds that the flaw “requires an x86-64 processor with AVX-512 VBMI support” and that “ARM64 CPUs are not affected”.

On processors, Calif lists Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD EPYC and Ryzen chips from Zen 4 onward. OpenJDK’s advisory lists affected versions as “27.0.1.0.3, 25.0.5.0.3, 21.0.13.0.3, 17.0.21.0.3, 11.0.33, 8u512, and earlier” for the release as a whole, and says “We recommend that you upgrade as soon as possible”.

How AI Was Used

Calif researcher Quang Luong writes that he “told Codex to look at those areas like any serious security researcher in 2026”, meaning OpenJDK’s JIT compiler, intrinsics, widely used APIs and cryptography. He writes that “most early results Codex brought me were slops”, and that “Codex brought me a finding for a memory corruption in the Base64 decoder itself”.

Calif says “it took our own Quang Luong an hour to discover the vulnerability, and two more hours to develop the first working exploit”. It calls the approach the “centaur hacker”: “AI brings the horsepower. Our researchers bring the context”. A Techmeme summary of a New York Times report says Calif used OpenAI’s GPT-6 Cyber; the Calif write-up we read names Codex.

The Business Lever

The lever is AI-assisted vulnerability research becoming cheap enough for a small firm to run against core infrastructure. The same firm appears in our report on Apple flaws found with GPT-6 Cyber. OpenAI President Greg Brockman said in an interview with The Information that there is no question cybersecurity will be a bigger business than coding.

Calif’s own recommendation is what it calls “vibe-uncoding”: using AI to “identify the parts a system actually needs, then remove everything else to reduce its attack surface”. It argues the Base64 speed-up “helps applications that decode enormous amounts of data” but “ships to everyone else too”.

The Structural Read

As we read Calif’s account, the cost of finding a critical bug in core infrastructure has dropped to an hour of a researcher’s time and under $200 of tokens, with two more hours to a first exploit.

As we read the two scores, the vendor and the finder differ on severity: OpenJDK gives 8.1 under CVSS 3.1, Calif 9.2 under CVSS 4.0, and the two systems weigh factors differently.

Calif, Java4Shell write-up

“The best hacker today is neither AI nor human.”

Three Implications

FOR JAVA OPERATORS Calif says all Java users should update immediately, or start OpenJDK with -XX:-UseBASE64Intrinsics or, on x86-64 only, -XX:UseAVX=2; it says ARM64 CPUs are not affected.

FOR VENDORS OF JAVA APPS Calif lists pre-auth RCE in Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate in default configurations, and says it is holding back the product-specific exploits for now.

FOR AI LABS Calif says Codex surfaced the Base64 finding after many unusable results; a Techmeme summary of the New York Times says the model was GPT-6 Cyber, which Calif’s write-up does not name.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework says: “Value capture migrates up from the model to the harness.”

As we read it, Calif’s account fits that pattern: it says it built pipelines and strategies first, then directed AI agents to use them, so the firm wrapping the model does part of the work. The framework predates this story and does not discuss Java or Calif.

Update, 17 November 2026, 22:40 UTC: Oracle’s own record widens the list to GraalVM; CISA has not listed it as exploited. Oracle’s entry for CVE-2026-114938 went up on the National Vulnerability Database at 22:18 UTC (NVD status: Received). Oracle says the flaw is in the Compiler component and names three products: Oracle Java SE (8u511b33-perf, 17.0.21.0.3, 21.0.13.0.3, 25.0.5.0.3, 27.0.1.0.3), Oracle GraalVM for JDK (17.0.21.0.3 and 21.0.13.0.3) and Oracle GraalVM Enterprise Edition (21.3.20.0.3). It rates it 8.1 under CVSS 3.1 and calls it a “Difficult to exploit vulnerability” that “allows unauthenticated attacker with network access via multiple protocols” to compromise those products, with “takeover” as the result of a successful attack. At our read at about 22:4x UTC, CISA’s Known Exploited Vulnerabilities catalog (last updated 17:53 UTC, 1,782 entries) did not include the CVE. Our count: 8 affected builds across 3 Oracle products, 3 more than the 5 Java SE builds alone; Oracle’s 8.1 (CVSS 3.1) sits below Calif’s 9.2 (CVSS 4.0), and the two scores use different scales.

What Is Not Established

We could not read the New York Times article, so the GPT-6 Cyber attribution rests on Techmeme’s summary of it; Calif’s write-up names Codex and does not name a model version. CVE-2026-114938 was not yet listed in the US National Vulnerability Database when we checked at 21:33 UTC. Calif’s product list and exploit claims are its own and have not been independently tested by us. We did not contact Calif, OpenAI or Oracle.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The Agentic Harness War

A Business Engineer analysis of where value pools as AI agents move from the model to the harness around it.

Read the Map of AI →

The Bottom Line

Calif says a researcher using Codex found Java4Shell, a remote-code-execution flaw in OpenJDK’s Base64 decoder, in an hour for under $200 of tokens, and built a first exploit in two more hours. OpenJDK rates it 8.1, the highest of 10 CVEs in its 17 November advisory, and Calif says users should update immediately or disable the affected intrinsic.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. The description of Java4Shell, the affected products, distributions and processors, the one-hour discovery, the two-hour exploit and the under-$200 token cost are as stated by Calif in its write-up of 17 November 2026, which we read in full. The 8.1 CVSS 3.1 score and affected release lines are from OpenJDK’s 17 November vulnerability advisory. The GPT-6 Cyber attribution is from Techmeme’s summary of a New York Times report we could not read. We did not contact Calif, OpenAI or Oracle.

Sources: NVD, CVE-2026-114938 (Oracle CNA record), read 17 Nov 2026 · CISA KEV catalog, read 17 Nov 2026 · Calif, Java4Shell: Write once, RCE everywhere, 17 Nov 2026 · OpenJDK Vulnerability Advisory: 2026/11/17 · Techmeme, summary of The New York Times, 17 Nov 2026

Scroll to Top
FourWeekMBA