AI-Found ‘Java4Shell’ Flaw Puts Banks, Hospitals’ Java Servers at Risk
A critical flaw in Java's Base64 decoder, found by security firm Calif with OpenAI tools, could allow takeovers of Keycloak, Jira and Jenkins servers.
A serious security flaw in Java, one of the world’s most widely used programming platforms, was uncovered with the help of artificial intelligence and could let attackers take control of servers at banks, hospitals, airlines and government agencies, according to several cybersecurity outlets. The bug, nicknamed “Java4Shell,” sits in the code Java uses to decode Base64 text, and it can be triggered with a string that begins with a long run of the letter “A.” Oracle has released fixes, and the U.S. Cybersecurity and Infrastructure Security Agency (CISA), the federal government’s main cyber defense agency, has urged organizations to patch right away, Engadget and ITPro reported.
What happened
The vulnerability was found by Calif, a cybersecurity research startup. Base64 is a common way of turning data into plain text characters so it can travel safely through web systems, for example in login headers and the digital tokens websites use to identify users. Engadget reported that feeding Java’s decoder a string of 64 “A” characters followed by an equals sign causes a memory error. On many machines that simply crashes the program, but when the error is manipulated carefully it can give an attacker unauthorized access to the system.
The French technology site IT-Connect described the mechanism in more detail: a misplaced “=” padding character leads the decoder to calculate a negative output size, so it writes up to 62 bytes past the memory area it is supposed to use. That can overwrite nearby memory inside the Java Virtual Machine, the software that runs Java programs. The flaw has been assigned the identifier CVE-2026-114938, according to IT-Connect.
The bug only affects a faster version of the decoder that uses a processor feature called AVX-512 VBMI. That feature is found in Intel Xeon server chips from the Ice Lake generation onward and in AMD processors since the Zen 4 generation, IT-Connect reported. ITPro noted that machines running on ARM64 chips are not affected.
The role of AI
Accounts differ on exactly which AI tool was used. Engadget and ITPro reported that Calif relied on GPT-6 Cyber, an OpenAI model that ITPro said launched on Nov. 5 for a select group of approved security clients and is not available to the public. IT-Connect, however, reported that Calif credited its researcher Quang Luong working with Codex, OpenAI’s coding agent, and described the GPT-6 Cyber link as unconfirmed.
Either way, the speed is striking. “It took our own Quang Luong an hour to discover the vulnerability,” Calif said in its announcement, as quoted by ITPro. The firm added that it needed “two more hours to develop the first working exploit,” and that the AI computing cost came to less than $200.
Why it matters
Java runs a huge share of corporate and government software. Engadget reported that about 90 percent of Fortune 500 companies use it. Calif said it was able to achieve remote code execution, meaning it could run its own commands on a target machine without logging in, against several popular business tools in their default settings, including:
- Keycloak, an open-source login system used by organizations such as CERN and NASA
- Atlassian Jira and Confluence, widely used project-tracking and collaboration tools
- Jenkins, a common tool for building and deploying software
- JFrog Artifactory and PingFederate, used to store software and manage sign-ins
Other products, such as Apache Kafka and Spinnaker, could be compromised only by attackers who were already logged in, according to ITPro and IT-Connect.
There are limits to the danger. Calif told reporters the bug is not easy to turn into a weapon, and IT-Connect noted that there is no single exploit that works everywhere; each application needs its own attack chain. ITPro reported it is not known whether the flaw has been used in real attacks, though the security firm Blackpoint Cyber said it had reproduced attacks that crashed a Java application and pulled secrets from its memory.
The severity ratings also differ. Calif scored the flaw 9.2 out of 10, while Oracle rated it 8.1, judging exploitation to be complex, IT-Connect reported.
What comes next
Oracle, which oversees Java, released its November security update on Nov. 17 with fixes for 655 flaws, including 11 in Java SE, according to IT-Connect. Engadget noted that Oracle’s update did not specifically mention this bug, and ITPro said the company has not otherwise commented. Patched versions are available for Java 17, 21, 25 and 27, and Amazon, Azul and Oracle’s GraalVM have issued their own fixes, while several other Java distributors had not yet done so as of Nov. 18, IT-Connect reported. Older Java 8 and Java 11 releases are generally not affected.
For organizations that cannot update quickly, ITPro said there are workarounds that switch off the vulnerable decoder. Even so, Calif warned that the bug could linger in unpatched systems for five to 10 years, Engadget reported, a reminder of how slowly fixes spread through older corporate software.


