ThinkFacility Sign in
  1. Home
  2. News
  3. OpenAI

Calif says one researcher using OpenAI's Codex found a critical Java flaw in an hour

The AI cost under $200, and Calif says it got into Jira and Jenkins without a password.

On November 17, 2026, the security firm Calif disclosed Java4Shell, a memory-corruption bug in the Base64 decoder of OpenJDK, which Calif calls the most popular Java runtime. Oracle shipped fixed builds the same afternoon.

Calif says the bug lets a remote attacker with no credentials overwrite memory on a server and reach remote code execution. Its researcher Quang Luong found it with Codex, OpenAI's coding agent, in about an hour.

Found by
Quang Luong of Calif.io, working with Codex
Time to find it
an hour, then two more hours for the first working exploit
Token cost
less than $200
Severity
9.2 out of 10 under CVSS 4.0, Calif's score
Patch
Oracle, 1:00pm PT on Nov 17, 2026

How Codex found it

Luong writes that he aimed the audit at the parts of Java that are hardest to get right: the JIT compiler, intrinsics, widely accessible APIs and cryptography. In his words, he "told Codex to look at those areas" the way a serious researcher would, and he didn't expect to find anything trivially exploitable.

Most of what came back was junk (he calls it slop), and he says he kept prompting for around 20 minutes before Codex flagged the decoder. The bug sits in an intrinsic that speeds up Base64 on newer x86-64 processors. A malformed string makes it calculate a negative output length, and it then writes 62 bytes past the end of an array.

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.

From Java4Shell: Write once, RCE everywhere

Calif's name for this way of working is the centaur hacker. It says the AI brings the horsepower and its researchers bring the context, and it used the same approach on PHP2Shell, the bug it disclosed before this one.

The best hacker today is neither AI nor human.

Java4Shell: Write once, RCE everywhere

Who has to patch

Calif says it got code running without a login on default installs of six products: Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate. On Kafka, Solr, Elasticsearch, Gerrit and Spinnaker it needed to be signed in first. It calls both lists non-exhaustive.

The flaw is in OpenJDK 17.0.6 and newer, so it's also in the builds Amazon, Microsoft, Red Hat, Azul and others make from it. It only fires on x86-64 chips with AVX-512 VBMI support, and ARM64 machines aren't affected.

Oracle's own advisory scores the bug lower than Calif does: 8.1 under CVSS version 3.1, with attack complexity rated high. It credits two people for it, Quang Luong of Calif.io and Albert Yang. The same update credits Nicholas Carlini of Anthropic for a different flaw.

All Java users should update immediately.

From Java4Shell: Write once, RCE everywhere

For a server that can't be updated today, Calif gives two startup flags. -XX:-UseBASE64Intrinsics turns off the fast decoder, and -XX:UseAVX=2 turns off AVX-512, though only on x86-64: OpenJDK on ARM won't start with it.

Why it matters outside the industry

Most people never see the Java underneath the services they use. Calif says it powers 90% of Fortune 500 companies, and it lists banking, aviation and telecommunications among the sectors that run on it. That's Calif's figure, and we haven't checked it against a second source.

The speed is the part we'd watch. On November 16 we covered Google's finding that attackers already use an AI tool that turns published patches into exploits with no human involved. The same day, Anthropic cut its disclosure deadline to 30 days, saying another AI would find a bug it sat on. Oracle's fix went public on November 17.

For a company running Jira or Jenkins on its own servers, that puts the update ahead of the next maintenance window. Calif is holding back its product-specific exploit chains until everyone has had a chance to upgrade, and says it will publish them later. It didn't give a date.

More on OpenAI

All OpenAI stories