OpenAI Unveils GPT-6 Cyber to Find and Exploit Software Flaws
OpenAI has introduced GPT-6 Cyber, a model for security work designed to detect, exploit and propose fixes for software vulnerabilities. The company cites impressive figures that cannot yet be independently verified.
OpenAI has introduced GPT-6 Cyber, a new AI model for security work that can not only find software vulnerabilities but also exploit them and suggest ways to fix them. The company is thereby targeting a field in which attack and defense lie closely together. The announcement first became known through a report by the portal The Decoder.
According to OpenAI, the model has uncovered 81 previously unknown vulnerabilities in the Firefox web browser. It is also said to have developed 948 working exploits for the Linux kernel, meaning programs that can actually take advantage of a vulnerability. The number of exploits generated is said to be significantly higher than the result of the predecessor model. Concrete comparison figures for that predecessor, however, are not given in the reporting. It therefore remains unclear how large the difference actually is.
According to the company, the model's strength lies in combining three tasks: finding a vulnerability, developing an exploit and proposing a fix. The same capability that helps a defender close a gap quickly can show an attacker how to exploit it. It is precisely this dual nature that makes models of this kind both interesting and risky for security professionals.
Crucially, all the figures cited are based on statements from OpenAI itself. According to the available information, no independent review has been conducted. The sourcing also rests on a single report from a specialist portal. Whether the vulnerabilities discovered have already been reported to the vendors or fixed is likewise unclear. It therefore remains uncertain how large the actual security gain is for the affected projects, namely the Firefox browser and the Linux kernel, to which the model was applied.
It is also unclear who will get access to GPT-6 Cyber. No details are available on the protective mechanisms intended to limit misuse in generating exploits either. This question is central for models of this kind, since the abilities to find and exploit gaps cannot be separated when a system is meant for defenders. Without clarity about access and control, the significance of the announcement can only be assessed to a limited extent.
Source: the-decoder.com
10
