TL;DR
Security start-up Calif has used AI to find a serious vulnerability in Java, the programming language behind systems at banks, hospitals and governments, BBC News reported on 18 November. The flaw, which Calif calls Java4Shell, sits in Java’s memory functions and could let an attacker run commands remotely. Calif did the work with OpenAI’s GPT-6 Cyber, which the public cannot yet use.
What was found
Java is supposed to make software safer by managing memory on the programmer’s behalf, which is why Professor Toby Walsh, an AI researcher, called the bug “such an unlikely and serious flaw”. He told the BBC it had been “hiding in plain sight for 12 years”.
Oracle, which owns Java, released a “critical patch update” on Tuesday, 17 November, containing security fixes, but has not confirmed this particular bug. The BBC said it was unclear whether any organisation had been affected, and that it had asked Oracle for comment. The worry is that attackers could now go looking for the same weakness in systems that have not been patched.
A second find in weeks
This is Calif’s second discovery of its kind. It previously found a flaw in PHP, the language behind sites such as Wikipedia. Professor Alan Woodward of the Surrey Centre for Cyber Security told the BBC that the PHP flaw affected about 30 million websites.
Walsh said AI is “now exposing more flaws much more quickly”, including some that people might never have found unaided. The model Calif used, GPT-6 Cyber, is offered by OpenAI to approved security researchers rather than the general public, and OpenAI’s own launch material for its new GPT-6.2 models lists PHP2Shell as an example of AI helping researchers uncover critical bugs.
The discovery arrived during a run of technology failures, including a large Microsoft outage. Marco Rubio, the US secretary of state, has said there is “zero evidence” that those incidents are connected.
Looking forward
For UK organisations, the practical step is the same as with PHP: check that Oracle’s latest Java update is applied and ask suppliers whether their products run affected versions. In our view, the bigger shift is in who finds flaws. When a small start-up with a restricted AI model can surface two decade-old bugs in widely used languages within weeks, defenders and attackers alike should expect the pace of disclosure to keep rising.