AI Helped Find a Critical Java Flaw in an Hour for Under $200

It took one researcher about an hour, working with an AI model, to find a critical flaw in Java, and two more hours to turn it into a working attack. The whole job used less than $200 of AI computing, according to Calif, the security start-up behind the find.

The firm disclosed the flaw on Tuesday. It calls it Java4Shell. Its official identifier is CVE-2026-114938. It scores 9.2 out of 10 on CVSS 4.0, the scale the security industry uses to rate how dangerous a bug is. Oracle and the OpenJDK project, which maintains the open-source core of Java, released fixes the same day.

The cost is the number to watch. Small teams using top AI models now build in days or hours the kind of advanced attacks that once took elite hackers at spy agencies months to construct, the New York Times reported. Java4Shell is the latest example, in software that runs a large part of the world’s business systems.

What Java4Shell is

Java is one of the most widely used programming platforms. The security firm says it powers 90% of Fortune 500 companies, and that banks, airlines, telecoms firms and national security systems depend on it.

The bug sits in code that decodes Base64. Base64 is a common way to turn binary data, such as uploaded files or login tokens, into plain text so it can travel safely over the internet. To decode it faster, Java uses special instructions on some Intel and AMD server chips.

The researchers found that a string of 64 letter A characters followed by one equals sign makes that fast code write 62 bytes past the end of the memory it was given. This is called memory corruption. Data spills into neighbouring memory that it should never touch. Java is designed to prevent exactly this kind of error. The flaw is in a small piece of hand-written, low-level code that Java relies on.

With careful work, an attacker can use that overflow to read sensitive data from a server’s memory. They can also use it to run their own commands on the server, which is known as remote code execution.

How far it reaches

The security firm says it achieved remote code execution without logging in on six widely used business tools in their default settings. They include Jenkins, Jira, Confluence and the Keycloak login system. It did the same on five more tools, among them Elasticsearch and Apache Kafka, after logging in. The firm says both lists are incomplete.

Ten Java distributions built on OpenJDK are affected, the firm says, including Amazon Corretto, Microsoft’s build of OpenJDK and Oracle Java SE. The flaw affects OpenJDK 17.0.6 and later versions.

There are limits. The flaw only works on x86-64 processors that support a set of instructions called AVX-512 VBMI. These include recent Intel Xeon chips and AMD chips from the Zen 4 generation onwards. ARM processors are not affected.

Datadog, a cloud monitoring company, tested the bug separately. Its researchers found that crashing a vulnerable application is far easier than taking it over, and that no single attack works against every application.

Thai Duong, the start-up’s chief executive, told the New York Times that the bug was not easy to turn into an attack, which gives defenders more time. “But vulnerabilities have a very long tail,” he said. He predicted the flaw could threaten systems that are not updated for the next five to 10 years.

How the firm used AI

The start-up says it used AI both to find the flaw and to build attacks with it. The New York Times reported that the model was GPT-6 Cyber from OpenAI, and the BBC says that model is not yet available to the public. The same report says the start-up worked with government agencies in five allied countries, and with large vendors, before it announced the flaw.

Quang Luong, a researcher at the firm, found the bug in about an hour and built the first working exploit in two more hours, it says. After its team set up the tools and strategy, most of the later attack chains for individual products were built “almost autonomously with AI.” The firm calls this approach the “centaur hacker”, part human and part AI.

The New York Times says it independently confirmed the discovery using technical details of the attack that the firm shared. Its test took less than an hour.

This is the start-up’s second major AI-assisted find in about a month. In October, AI models helped uncover PHP2Shell, flaws in the PHP engine behind roughly 70% of websites. The disclosure plan for those flaws broke down when details leaked before the planned release date.

What the numbers say about defenders

OpenJDK says it publishes security fixes four times a year, on a fixed schedule. The November fix is its eighth security advisory of 2026. Not every extra release this year is linked to AI-found bugs. Even so, 2026 has already seen twice the planned number.

CISA, the US cyber defence agency, told organisations to patch immediately. Its alert also tells them to look for copies of the runtime bundled inside other software, containers and appliances, which may not show up in software inventories.

“We have seen this taking shape for some time, with AI being used to discover a growing number of latent vulnerabilities across software and the digital services we all rely on,” Ollie Whitehouse told the New York Times. He is chief technology officer of Britain’s National Cyber Security Centre.

Duong expects AI to build the most advanced attacks on its own at some point. “I don’t know how much time we have — maybe three months, maybe three years,” he said. “But right now, the best hacker is the centaur hacker,” he added.

What to do now

The main risk sits on servers run by companies, banks and public bodies. But the firm says a Java program that connects to a malicious server can also be attacked, so anyone with Java installed should accept updates when they are offered.

Anyone who runs Java on servers should update now. The first fixed releases are 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. Where an update is not possible yet, the security firm and Datadog list two startup settings that switch off the vulnerable code. Java processes must be restarted for either step to take effect. The start-up says it will keep its attack code for individual products private until more systems are patched.

Why it matters: A critical bug in software used by most large companies was found and turned into a working attack in about three hours, for less than $200 of AI time. Fixes are out, but the start-up’s chief executive expects unpatched systems to stay exposed for years.

The post AI Helped Find a Critical Java Flaw in an Hour for Under $200 appeared first on DailyAI.