Beating AI News Flash: Cybersecurity company Calif, with the help of OpenAI's GPT-6 Cyber, discovered a critical vulnerability in Java's low-level code named Java4Shell. Researchers said it took only 1 hour to find the vulnerability and another 2 hours to develop the first working exploit, with AI token costs under $200.
Java is used by about 90% of Fortune 500 companies worldwide, running extensively in the backend systems of banks, governments, hospitals, and major internet companies. Calif has verified in vulnerability-eligible environments that enterprise software such as Keycloak, Jira, Confluence, and Jenkins can be remotely compromised, many of which require no login. Keycloak in particular handles enterprise identity authentication and login management; once compromised, it could also endanger other business systems that depend on it. After obtaining the attack verification materials, The New York Times also independently reproduced the attack in under 1 hour.
The vulnerability is hidden in Java's Base64 decoder. This feature is widely used to process login credentials, file uploads, and network communication data. To improve speed, Java uses low-level assembly code to process this data, but this left a memory out-of-bounds write vulnerability. Attackers can overwrite other memory through specially crafted input, thereby stealing sensitive data or even taking control of servers. Java was originally known for strong memory safety, and the problem this time lies precisely in the low-level acceleration code that bypasses conventional memory checks.
Calif just used AI last month to discover a critical vulnerability in PHP, and this time it has found a low-level flaw in Java. Before disclosure, the team had coordinated remediation with relevant government departments of the United States, the United Kingdom, and other Five Eyes member countries, as well as tech companies such as Google and Microsoft. The U.S. Cybersecurity and Infrastructure Security Agency (CISA) requires enterprises to install patches immediately, and Google has confirmed that it has deployed protections and fixes across its services.
South Korea's Presidential Office National Security Office convened an emergency meeting on November 17 with government, military, and financial regulatory authorities. According to South Korean government statistics, of the country's 15,618 public information systems, 11,132 use Java, accounting for 71.28%. Relevant departments urgently investigated Java usage and coordinated security updates. Oracle released a patch the same day.
At present, attacks still have barriers: they require a specific version of Java and processors supporting AVX-512 VBMI. Different software also requires separately developed attack methods. As of November 17, Cisco's security team had not observed actual attack activity.
The Calif team refers to this approach of having top security experts guide AI to discover and exploit vulnerabilities as the "centaur hacker." CEO Thai Duong told The New York Times that in the future, AI may independently carry out the most complex cyberattacks.
"I don't know how much time we have left — maybe three months, maybe three years. But the strongest hackers right now are centaur hackers."
