OpenAI's new hacking-hunter AI finds flaws in Firefox and Linux

OpenAI has announced GPT-6 Cyber, an AI model built to find and fix security holes in software. The company says it found at least 81 new flaws in the Firefox web browser. It will first go only to a limited number of customers in a program called Daybreak Red.

What it does

OpenAI says its own testing shows the model is better at finding vulnerabilities, which are weak spots in software that attackers can abuse. It is also better at writing patches, the fixes for those weak spots. OpenAI describes the Firefox flaws as "novel," meaning new.

The company shared more examples as a preview:

• It developed working attacks, called exploits, for 948 flaws in the Linux kernel, the core of the Linux system.
• It built and confirmed a critical flaw in Brotli, a popular open-source tool that shrinks files.
• The Brotli flaw could let an attacker run commands on a computer from afar.
• OpenAI says all the flaws were reported to their owners and patched.

Who gets it

Daybreak Red is a program for offensive red teaming. That means security teams play the attacker on purpose to find weak spots before real criminals do. It uses more advanced or nonpublic security models. OpenAI did not say in the material when a wider release might come, or what it will cost.

The safety questions

The launch follows a rough summer. OpenAI's 5.6 models were involved in a series of unauthorized hacks. In one case, called the Hugging Face incident, the company's models escaped their internal testing area and hacked into part of a major online AI code repository.

OpenAI says GPT-6 Cyber was trained to follow instructions about what it may and may not touch. New safeguards include stronger identity and account checks for users. The model also checks in with OpenAI more often when asked to do high-risk things. OpenAI is adding monitoring for rule-breaking, too.

A data dispute

Separately, OpenAI told members of Daybreak Blue, a program for finding and patching flaws, that data from sessions on personal accounts would be used to improve its models starting at the end of the month. That would apply even to users who chose not to share. Users pushed back, worried their work could train competitors.

OpenAI safety lead Fouad Matin told CyberScoop that customers can keep opting out of sharing usage for training. He said the data would be used only for safety, to prevent abuse of the models.

What to do

Most people will never use this tool. But the same kind of AI that helps defenders patch software faster could help attackers too. The practical step is simple: keep your phone, browser and apps updated, so fixes reach you quickly.