What A Time To Be Alive
Choose language: English

Researcher uses AI to find serious flaw in Java

A flaw in Java can, under certain conditions, let attackers run their own code on servers, sometimes without logging in. A researcher found it in an hour with OpenAI's Codex. Oracle and other makers released updates .

A stream of white tiles falls into one compartment of a wooden sorting tray, piles up high and slides over the edge into the next compartment, where a bunch of brass keys is half buried under the tiles.
Data spilling over the edge of its own compartment: that is how the Java flaw works, which Oracle and other makers are now fixing (illustrative photo, generated with AI)

Java is one of the most widely used programming languages for servers: many login systems, online shops and business applications run on it. , the security company Calif disclosed a serious flaw in OpenJDK, the open foundation on which almost all versions of Java are built. The flaw is called Java4Shell (officially CVE-2026-114938). Oracle, which manages Java, released a security update the same day.

What is wrong

The fault sits in the part of Java that decodes Base64. Base64 is a way of sending data, such as a file or login details, as ordinary letters and numbers. On newer server processors, Java uses a special shortcut to do this faster. That shortcut contains a calculation error: with a short, specially crafted piece of text, Java writes data outside the place in memory that is meant for it. An attacker can use this to read secret data from memory, such as keys and passwords, or to make the server carry out their own commands.

The flaw only affects servers with certain processors: Intel Xeon processors from the Ice Lake generation onwards and AMD processors from the Zen 4 generation onwards. Servers with ARM-type chips are not affected. OpenJDK 17.0.6 and newer are affected, as are the Java versions built on it, including those from Oracle, Amazon, Microsoft, Red Hat and Azul. A program is only vulnerable if it processes Base64 data from outside, and an attack has to be adapted to each program.

Calif showed that the attack works on well-known software with its default settings. With the login system Keycloak, the collaboration tools Jira and Confluence and the build server Jenkins, among others, it worked without logging in. With the search software Elasticsearch, among others, it worked after logging in. Calif gives the flaw a severity score of 9.2 out of 10; Oracle, using an older scoring method, arrives at 8.1.

Found with AI

Calif researcher Quang Luong had Codex, OpenAI's coding assistant, search the parts of Java where this kind of fault is often found. According to Calif, Codex found the fault in the Base64 code within an hour. A first working attack took another two hours, and the AI use cost less than 200 dollars in total. According to Calif, AI built most of the later attacks on individual programs almost on its own. Calif does not say which version of the model was behind Codex or exactly when the search took place; the flaw was made public .

What administrators are doing now

, the security company GreyNoise saw the first attempts to exploit the flaw, but no successful attack against its own sensors yet. By then Oracle, Amazon and Azul had already released fixed versions, and the Eclipse project had started work on one. Red Hat and Ubuntu had no update yet , and Debian had only updated its development version.

Anyone who cannot update yet can temporarily switch off the Base64 shortcut by starting Java with the option -XX:-UseBASE64Intrinsics. Updating remains the best solution. It is also wise to check software packages that bring their own copy of Java.

What does this mean for the road to the singularity?

That AI helps to find faults in software is not new. A day later, two flaws in the NGINX web server were also fixed; the security company Wiz had found them with AI, and one of them had been in the code since 2015. What is new with Java4Shell is how little it cost: a serious fault in one of the most widely used building blocks of server software came to light in an hour of work, for less than 200 dollars. If searching for flaws becomes this cheap, anyone can do it, defenders and attackers alike. A software maker that does not have its own code checked by such an assistant is in effect leaving that work to an attacker.

If that trend continues, the work shifts from finding to fixing. An AI that finds a flaw in an hour can also help to repair it and test the repair. With Java4Shell, the update was already available on the day the flaw became known. If every new version of important software is searched this way in advance, many faults will come to light before they ever reach users. The last step remains human work: actually getting the update onto every server. Those who organise that well will gain the most from a time in which faults are found faster than ever.

What is the technological singularity? Read the explanation →

Sources

  1. Tom's Hardware
  2. BleepingComputer
  3. Tenable®
  4. Calif: Java4Shell