AI

Java4Shell: AI found a hole in Java in an hour, and an exploit took two more

Java code on a screen with a symbolic padlock being broken: a software vulnerability

Photo: Daniel Trylski / Pexels

Researchers at Calif have found a vulnerability in Java, which they call Java4Shell and describe as critical. The search took about an hour. A working exploit appeared two hours later, and the AI cost less than $200 (er10.kz). The flaw could potentially give attackers access to sensitive data or let them run malicious code on other people’s computers.

Three hours and under $200: how Calif found Java4Shell

Calif is a cybersecurity startup, according to TNW. As er10.kz reports, its team used GPT-6, OpenAI’s specialized model for security work. The New York Times reported that it was GPT-6 Cyber specifically, and ran the attack itself in a test that took less than an hour (thenextweb.com).

Calif says researcher Quang Luong pointed OpenAI’s Codex agent at the Java internals. The first working attack was put together in two more hours. By the company’s estimate, all the AI tokens together cost less than $200.

An exploit is a ready-made way to take advantage of a flaw.

That adds up to three hours of work. We have already covered how OpenAI pitches GPT-6 Cyber to security teams. Here, the model was used to hunt for bugs in Java.

64 letter A’s and 62 extra bytes

The bug sits in the fast version of Java’s Base64 decoder. Feed it 64 “A” characters and one “=”, and the decoder writes 62 bytes outside its own memory. Java rejects that input only after the overflow has happened (thenextweb.com).

Not every computer is affected. Only Intel and AMD processors with the AVX-512 VBMI instruction set are vulnerable. ARM chips are not.

Calif claims that code execution without authentication works on default settings in Keycloak, Atlassian Jira and Confluence, JFrog Artifactory, Jenkins and PingFederate. The company says the list is incomplete.

Keycloak is an authentication system used by large companies and government organizations. As er10.kz writes, the possible impact on Keycloak in particular raised special alarm.

Calif sizes up the scale this way: 90% of Fortune 500 companies run on Java. And er10.kz points out that Java programs run banks, hospitals and other infrastructure.

Attacks are already being tried, but there is no universal exploit

Calif disclosed the vulnerability on Tuesday. The same day, Oracle and the OpenJDK project released fixes (thenextweb.com). The US agency CISA urged organizations to update immediately.

Independent companies are tracking exploitation attempts. Tenable (tenable.com) writes that as of November 18, GreyNoise had recorded attack attempts, but no confirmed successful ones. VulnCheck says a working exploit is available to its customers. The flaw is not in CISA’s Known Exploited Vulnerabilities (KEV) catalog.

watchTowr has also seen attack attempts on its sensors. Other labs reproduced crashes and memory leaks, but not code execution so far.

There is no universal exploit: each application needs its own. Calif is holding back its product-specific exploits until users have updated.

A leak before the patch and a charge of “publicity-maxxing”

There is an awkward detail in this story. The condition that triggers the bug appeared publicly as early as September 9, in a proposed fix for OpenJDK. Its author later closed it without merging. Calif wrote on X that news of the vulnerability leaked before Oracle’s patch came out (thenextweb.com).

The company also has a critic. Jan Schaumann of Akamai accused Calif of “publicity-maxxing” on the Openwall mailing list.

Calif’s previous big find was the PHP2Shell vulnerabilities in October. According to er10.kz, the researchers had also reported earlier that they used AI to find serious security problems in PHP, which powers millions of websites.

“Centaur hackers” and the flip side

Experts call what’s happening the start of the era of “centaur hackers”: a human and an AI working together. They warn that in the future AI may hunt for vulnerabilities and carry out attacks on its own, with almost no human involved.

The same tools can help developers find and close dangerous bugs faster, before criminals exploit them. Where OpenAI itself draws the line on risk, we covered in a piece on the threshold the company once called a red line.

Понравился материал? Поделитесь с другими:

Picture of Раимбек Искендиров
Раимбек Искендиров
Picture of Раимбек Искендиров
Раимбек Искендиров

Похожие материалы

A loan agreement with the annual effective rate line highlighted and a calculator beside it.

Reviews

Insurance and collateral valuation costs will count toward APR on loans in Kazakhstan from November 30

From November 30, Kazakhstan changes how APR is calculated and disclosed. See which costs ...
A loan agreement and a calculator lying next to an income statement.

Reviews

National Bank puts off its 8-years-of-income debt cap

The National Bank has delayed its cap on debt at eight years of income. ...
Close-up of cloud server racks, the infrastructure startups received AI credits to use.

Reviews

Alibaba Cloud awards $40,000 in AI credits to startups in Kazakhstan and Uzbekistan

Alibaba Cloud has awarded $40,000 in AI credits to 20 startups in Kazakhstan and ...

Читайте нас в Telegram и Instagram

Подписывайтесь на наши каналы, чтобы первыми получать самые свежие новости казахстанской IT-индустрии, анонсы грантов, стажировок и эксклюзивные интервью.