Synthetic Intelligence & Machine Studying
,
Governance & Threat Administration
,
Subsequent-Era Applied sciences & Safe Improvement
Calif Says OpenAI’s Codex Uncovered Reminiscence-Corruption Bug; Oracle Points Patch

Simply weeks after the invention of PHP2Shell, synthetic intelligence fashions discovered one other important vulnerability, this time in Java.
See Additionally: Introducing Unit 42 Frontier AI Protection: A Defender’s Information to the Frontier AI Impression
Cybersecurity startup Calif stated in a weblog submit Tuesday that it found CVE-2026-114938,a memory-corruption vulnerability within the Base64 decoder of OpenJDK, a well-liked Java runtime platform. It calls the flaw Java4Shell.
Since Calif’s disclosure, Oracle stated it launched a patch for the vulnerability.
Java is among the many hottest programming languages on the web, powering greater than 90% of the world’s largest corporations, together with many authorities methods.
Calif stated the vulnerability can reveal delicate reminiscence and permit for distant code execution. It added that it could possibly set off the corruption “by simply 64 A characters adopted by a single =.”
In discovering the flaw, Calif researchers solely spent $200 in tokens, and it took three hours – one to find the vulnerability and two to create an exploit.
“We found a bug in that optimized machine-level code that lets malformed enter overwrite neighboring reminiscence,” the corporate stated.
Calif discovered the vulnerability after one researcher, Quang Luoung, audited OpenJDK and selected to take a look at the platform’s compiler, cryptography and APIs. Luoung directed Codex, OpenAI’s coding agent, to take a look at every goal “like several severe safety researcher in 2026.” The corporate didn’t disclose which OpenAI mannequin ran on Codex throughout its investigation, however The New York Occasions reported it was the lately launched GPT-6 Cyber.
Whereas Luoung thought-about a big chunk of Codex’s responses to be “slops,” it discovered reminiscence corruption in one in all Base64’s decoders.
Calif stated one of many promoting factors of Java is its reminiscence security, however as a result of Java is written utilizing languages like C and C++, reminiscence corruption persists. The corporate stated the bug “sits within the hand-rolled meeting code that handles the ultimate few characters” of the exploit.
Calif stated there is no such thing as a common exploit for Java4Shell, noting that it depends upon how the group makes use of Base64.
“Calif was in a position to speed up the exploit improvement by spending preliminary effort constructing pipelines and methods, then directing AI brokers to make use of these for preliminary exploitation. Afterward, a lot of the later chains are constructed nearly autonomously with AI,” the corporate stated.
British cybersecurity knowledgeable Kevin Beaumont suggested his social media followers to remain calm. “There is a bunch of TLP RED bonkers briefing a few world ending Java vuln,” he wrote. “It isn’t world ending. It isn’t very thrilling. Simply patch as you usually do in a while at this time – as you must have final month as there was a special unauth RCE vuln then.”
Java4Shell is the latest important vulnerability discovered round a language that many web sites depend on. PHP2Shell, disclosed on Oct. 21, impacted the extremely popular PHP, which powers not simply web sites however the internet hosting platforms that help them.
AI mannequin suppliers and cybersecurity corporations have urged organizations to make use of AI brokers to defend in opposition to menace actors, particularly as AI fashions and brokers more and more reveal vulnerabilities and make it simpler for hackers to assault methods.








