Confirmed 17 Nov 2026 · Security A one-minute read

An AI audit found a critical flaw in Java in 1 hour, for under $200.

Calif Research used OpenAI's Codex to audit OpenJDK and found an unauthenticated memory corruption bug rated 9.2. Caveat: it only triggers on x86-64 chips with AVX-512, and ARM64 is unaffected.

Until now

AI helps with code review. Finding real zero days still takes months.

Now

An hour and $200 of tokens is the price of a critical bug.

Why it matters

Java runs behind most large companies, and the bug sits in its Base64 decoder, which almost everything calls. Calif reports pre-authentication code execution on Keycloak, Jira, Confluence, Jenkins, Artifactory and PingFederate. The firm says AI built most of the later exploit chains almost on its own.

Established

  • Calif used Codex; discovery took 1 hour, first exploit 2 more.
  • Token cost was under $200. CVE-2026-114938 is rated 9.2 critical.
  • OpenJDK advised and Oracle patched at 1:00pm Pacific Time November 17.

Not established

  • Exploitation in the wild. None was reported at disclosure.
  • That AI alone found it. A human directed the audit.
  • The Fortune 500 reach. That share is Calif's own estimate.
The five-minute readThe record, the timeline, the weight and 5 sources

The record

Calif Research's write-up at calif.io dates the OpenJDK security advisory and the Oracle patch to 1:00pm Pacific Time on November 17, 2026, which is 21:00 Universal Time and inside Edition 100's window. We swept security that run and left it empty, so this card is filed late under its true event date. The flaw is CVE-2026-114938, nicknamed Java4Shell, a memory corruption bug in OpenJDK's Base64 decoder. Calif scores it 9.2 of 10 on CVSS 4.0, which is critical. The trigger is 64 "A" characters followed by a single "=" padding character, 65 bytes, causing a 64-byte write into a buffer sized for 50; on OpenJDK 17 the trigger is 128 "A" characters and an "=". It affects x86-64 processors with AVX-512 VBMI support, and ARM64 is not affected. Calif's own words on method: "As with PHP2Shell, Calif used AI to find and weaponize Java4Shell"; "It took our own Quang Luong an hour to discover the vulnerability, and two more hours to develop the first working exploit"; "The total token cost was less than $200"; and "Afterwards, most of the later chains are built almost autonomously with AI". The firm calls the result "the centaur hacker, part human, part AI". CyberScoop, published November 17 at 22:24 Universal Time, reports that Luong audited OpenJDK using OpenAI's Codex. Calif reports pre-authentication remote code execution on Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate, and post-authentication execution on Apache Kafka, Apache Solr, Elasticsearch, Gerrit and Spinnaker, with a working break-in demonstrated against Keycloak. Product-specific exploit details are withheld until users upgrade. Calif says Java powers 90% of Fortune 500 companies. Oracle shipped patched builds in a November update covering 652 flaws across its product line. CISA issued an alert the same evening urging immediate patching; cisa.gov refused us, so we take the alert from The Record and CyberScoop. No exploitation in the wild was reported at disclosure, and the CVE was not on CISA's Known Exploited Vulnerabilities list.

How it got here

  1. Oct 2026

    Calif publishes PHP2Shell, which the firm describes as its first major AI-assisted find, using the same method.

  2. Nov 13

    Google Project Zero drops its patch grace period, saying models have shortened reverse engineering enough that zero patch-to-exploitation time is a reasonable model of attacker behaviour. Edition 100's Undercurrents.

  3. Nov 16

    Google Threat Intelligence Group reports attack software that turns security patches into working exploits with no human involved. Edition 097's card.

  4. Nov 17

    Oracle ships patched Java builds in a release covering 652 flaws and OpenJDK publishes its advisory at 1:00pm Pacific Time. Calif publishes the write-up and CISA issues an alert the same evening.

  5. Nov 18

    Coverage spreads through the security press. The CVE is still not on CISA's Known Exploited Vulnerabilities list, and no exploitation in the wild is reported.

Who has to react

Every operator of Keycloak, Jira, Confluence, Jenkins, Artifactory and PingFederate on AVX-512 hardware; Oracle and the OpenJDK maintainers; CISA, which urged patching without a KEV listing; OpenAI, whose Codex did the audit; and every security team that still budgets vulnerability research by headcount.

What to watch

  • Would strengthen itCalif or another party publishes a third AI-assisted vulnerability scored above 9.0 in widely deployed software, with a stated discovery cost under $1,000, by June 30, 2027.
  • Would weaken itNo AI-assisted vulnerability scored above 9.0 in widely deployed software is published by any party before September 30, 2027.
  • Would strengthen itCVE-2026-114938 is added to CISA's Known Exploited Vulnerabilities catalogue, or a breach is publicly attributed to it, by March 31, 2027.

Why it weighs this much

Signal strength
CostLow
CommitmentMedium
ReversibilityHard
PersistenceStructural

Cost is low, and that is the finding rather than a weakness in it: 3 hours of human time and under $200 of tokens. Commitment is medium. Calif put its name on a numbered CVE with a severity score and a product list, and OpenJDK and Oracle shipped patches the same day, which no party can quietly walk back. Reversibility is hard, because the bug, the trigger and the method are now all public. Persistence is structural rather than dated: the patch closes 1 decoder, while the price of discovery this demonstrates does not revert.

The words

Base64
A way of writing binary data using only letters, digits and a few symbols. Web software decodes it constantly, for tokens, headers and attachments.
Pre-authentication RCE
Running your own code on a server without logging in first. The most valuable class of bug, because there is no password for the attacker to steal or guess.
AVX-512
A set of fast instructions on some server chips. OpenJDK uses them to speed up decoding, which is where this bug lives.
CVSS
A 0 to 10 severity score for vulnerabilities, where above 9 is critical. It is assigned by the reporter or the vendor, and the two often disagree.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.

8 contributions. Thank you.

You have built
a record here.

Keep it: create an account to verify your contributions and put your name to the answers you choose. Your record stays yours.

Keep my record ↗
Optional: help show whose perspectives are missing
Kept with your answers under a random id, never with your name. Used for one thing: showing you, on your record, where you stand next to people like you, once at least 20 of them have answered.

Add your signature

Stand behind your judgment.

Create an account to attach your name to this answer. Your other answers show no name unless you sign them too.

Sign up to add your name ↗

Already have an account?