Critical Java flaw that AI found in an hour puts millions of corporate accounts at risk
How a $200 AI-assisted hack put cybersecurity pros on high alert.

Image by Cybernews
- CISA urged organizations to patch a critical Java flaw affecting OpenJDK and widely used enterprise applications.
- Researchers showed attackers could run commands remotely in Jira, Confluence, Jenkins, Keycloak, Artifactory, and PingFederate.
- Calif said AI helped find the flaw in about one hour and build an exploit in two more.
- Such Fortune 500 companies as Walmart, Amazon, JPMorgan Chase, Target, T-Mobile, and Nike, among others, are potentially affected by the flaw.
- Attackers may exploit vulnerable servers without logging in, especially after gaining access through stolen company VPN credentials.
Key Takeaways by nexos.ai, reviewed by Cybernews staff.
CISA has urged organizations to immediately patch the recently disclosed Oracle Java vulnerability that affects most of the Fortune 500 companies.
A critical flaw dubbed Java4Shell affects OpenJDK, the open-source platform underpinning Java applications worldwide. The researchers who discovered the flaw demonstrated its execution in six platforms.
Such Fortune 500 companies as Walmart, Amazon, JPMorgan Chase, Target, T-Mobile, and Nike, among others, are potentially affected by the flaw.
By successfully exploiting the vulnerability, hackers could potentially gain access to these companies' accounts.
Oracle released security fixes on November 17th, 2026, while CISA warned companies to immediately identify affected Java installations, apply patches as soon as they are available, review the Oracle security advisory, and monitor for indicators of suspicious activity.
What is Java4Shell?
CVE-2026-114938 vulnerability affects the Java Virtual Machine's (JVM) optimized Base64 decoder, which processes encoded data used by applications to handle authentication credentials, security tokens and other sensitive data.
The vulnerability lies within how Java processes specially crafted Base64 strings on supported x86-64 processors. When the decoder encounters malformed input containing an unexpected number of trailing padding characters, it incorrectly calculates how much memory it needs to store the decoded data.
Calif researchers discovered that a specially crafted malformed string could make the decoder write beyond its allocated memory, corrupting neighboring data.
Attackers can exploit it for remote code execution (RCE) and run malicious commands on the affected server.
Stay updated with our latest stories and follow us on social media
Be the first to discover new stories, ideas, and updates from our team.
Java4Shell is particularly concerning because attackers may not need to authenticate themselves before exploiting vulnerable applications. This happens because applications can process Base64-encoded data before checking whether a user is authenticated.
Attackers may exploit it to deliver malicious input through HTTP Basic authentication credentials, SAML artifacts, CSRF tokens, and MIME-encoded messages without first logging into the application.
The vulnerable code runs on certain x86-64 Intel and AMD processors supporting AVX-512 VBMI instructions. It affects Java versions starting from 17.0.6 and some releases of Java 21, 25, and 27.
Affected applications
Calif demonstrated remote code execution in six applications, including Atlassian Jira, Atlassian Confluence, Jenkins, Keycloak, JFrog Artifactory, and PingFederate.
The researchers also demonstrated attacks requiring authentication against Apache Kafka Broker, Apache Solr, Elasticsearch, Gerrit, and Spinnaker.
A successful attack could allow hackers to execute their own commands on the affected server, potentially steal sensitive information, access credentials, or establish a foothold for further attacks.
AI helped to build exploits in hours
Calif said its researchers used OpenAI's Codex to discover the vulnerability in approximately one hour and develop a working exploit in another two hours. The total cost for the exploit was less than $200.
Researchers explained that while AI agents helped automate some parts of developing an exploit, human expertise was still needed to direct the work and develop strategies.
Calif has withheld its exploit code while organizations apply security updates.
Other security researchers have independently reproduced the memory corruption, although they have not yet replicated Calif's remote code execution results.
Cybersecurity firm Wiz Atlas has said that they created an exploit in two hours using AI. “We should expect adversaries to target commonplace Java applications, where specific builds and memory layouts repeat across targets,” Wiz Atlas researchers said.
Motivated adversaries may tailor exploits with AI for specific targets, including home-grown applications,they added.
Blackpoint Cyber cybersecurity firm managed to crash a test Spring Boot application by sending approximately 28,300 malformed login requests over 20 seconds. Its researchers also demonstrated that the memory corruption could expose nearby secrets, but did not reproduce remote code execution.
A scenario: how can the Jira environment be hijacked?
Jira is often deployed on internal corporate networks rather than exposed directly to the public internet. In many organizations, employees and contractors can access it only through company VPNs.
But this does not necessarily keep it out of attackers' reach. Java4Shell does not bypass VPN authentication by itself. However, VPN credentials may be stolen.Ransomware groups often target corporate VPN credentials, granting them access to internal networks and systems that would otherwise be inaccessible from the internet.
Java4Shell could make the next step considerably easier. Once inside a company's network, an attacker with access to a vulnerable Jira instance could potentially exploit it to execute malicious code on the server without needing a Jira account.
Instead of spending time searching for weaknesses, obtaining application credentials, or figuring out how to move deeper into the network, attackers could have a direct route to establishing a foothold on an enterprise server.
The same risks extend beyond Jira. Other widely used applications, Confluence, Jenkins, and Keycloak, which were mentioned in the proof-of-concept exploits, are often intertwined with sensitive corporate infrastructure and internal project data.
For ransomware operators, this could significantly simplify the process of compromising corporate systems.