Calif said it had used a new OpenAI model to find a critical vulnerability in Java, the programming language behind a significant share of internet services and financial systems. It took a researcher an hour to find the flaw and another two to prepare an exploit; AI tokens cost less than $200. The Java4Shell vulnerability can give an attacker access to a system, expose sensitive data or allow remote code execution. Calif estimates that the issue could affect nearly every large company: about 90% of Fortune 500 companies use Java.
AI finds a Java vulnerability that could affect major companies
How Java4Shell works
The vulnerability was found in the Base64 decoder, a Java component that represents binary data as readable text. The error occurs when a program tries to process a 64-character string of “A”s followed by a single equals sign. As a result, Java mishandles the code, usually causing the program to crash because of memory overload.
Using OpenAI’s new GPT-6 cyber model, the researchers were able to control the memory corruption. This opens a path to system access: an attacker could obtain confidential data or run code remotely.
A separate risk involves Keycloak, a popular open-source authentication system used by large organizations, including NASA and UnitedHealthcare. Calif senior researcher Bruce Dang told The New York Times that Java is the foundation of the modern economy and that the vulnerability will affect nearly every large company.
Outages in digital systems
The discovery comes amid a series of disruptions to digital infrastructure. In recent weeks, banks and widely used Microsoft products went offline, a national air traffic control system failed, and Apple services experienced outages.
This week, hackers broke into the websites of hundreds of large organizations and injected malicious JavaScript to steal money and visitors’ bank card data. In Europe, freezer control systems used by grocery stores and blood banks were also hacked. No link between these incidents and AI-powered attacks has been established, but they heighten concerns about the reliability of digital systems that everyday services depend on.
Last month, Calif researchers used an OpenAI model to find serious vulnerabilities in PHP. The software underpins nearly 70% of websites on the internet.
“Centaur hacker”
Calif researchers call the emerging threat a “centaur hacker” — a combination of human and AI. For now, a specialist directs the model. But the researchers warn that a point may come when an autonomous model can cause damage without human control.
Dang told The New York Times that he does not know how much time remains before that happens: perhaps three months or three years. For now, he said, the best hacker is a centaur hacker.
In Dataist’s view, the Java case shows how AI is reducing the time and cost of finding and exploiting vulnerabilities in software that large organizations depend on. The source does not link the listed outages and attacks to AI, however: the concern is growing risk to infrastructure, not an established common cause behind the incidents.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X