- cyberscoop.com

OpenAI’s GPT-6 Cyber Campaign Finds Hundreds of Linux Kernel Exploits

In a mid-October campaign, OpenAI’s GPT-6 Cyber model generated 948 working Linux kernel privilege-escalation exploits at an average API cost of $1,279 each, with 588 targeting previously unknown flaws; OpenAI sent Linux maintainers 304 critical findings with proposed patches. The U.K. National Cyber Security Centre warned that AI tools are accelerating vulnerability discovery, requiring organizations to adapt patching and testing. OpenAI introduced GPT-6 Cyber on Nov. 5 for approved customers in its Daybreak Red program, and the model also identified 81 previously unknown high-severity Firefox flaws, with Mozilla beginning fixes. Dragos tested the model in a mock power plant, where it uncovered two unknown flaws in engineering software, prompting OpenAI to expand access to critical-infrastructure defenders while maintaining safeguards like IP allowlisting and credential-compromise monitoring.

I. Notice they buried the real story behind a mountain of numbers. 948 working Linux kernel privilege-escalation exploits at $1,279 each, 588 previously unknown, and a 30-hour run that cost over $1.2 million in API credits. That is not a research project — that is a production line for cyber weapons. They frame it as defensive, but ask yourself: who are the "approved customers" in Daybreak Red? What agency or military unit gets first access to every flaw the model finds? OpenAI controls the gate. They decide which vulnerabilities get disclosed, which get patched, and which get stockpiled for "authorized" use. The entire architecture of consent here is to normalize the privatization of zero-day discovery and to position a single corporation as the arbiter of critical-infrastructure security. That is not transparency. That is a monopoly on the means of digital coercion.

II. Look at what they hide between the lines. OpenAI "declined to detail" the cases where the model crossed its boundary in 0.07% of tests. A failure rate that low is irrelevant — what matters is what those boundary breaches revealed. In a mock power plant the model found flaws that could expose protective relay logic and circuit breaker trip settings. That is not a bug hunt. That is a targeting survey. Combine it with the fact that the U.K. National Cyber Security Centre is already telling organizations to "adapt their patching and testing processes" — you are being prepared for a world where these tools are used not only by defenders but by the same network that controls the model. The same elite institutions that fund OpenAI sit on the boards of the foundations that shape cybersecurity policy. It is the same revolving door. The pattern never changes.

III. Let me leave you with a question they do not want you to ask. Why did OpenAI include a "dedicated access group for approved critical-infrastructure defenders" in this announcement? Who qualifies as "approved"? The power grid, the water systems, the financial networks — the very sinews of civilization are being tested by an AI that only a handful of people are allowed to use. Every patch they release is a breadcrumb. Every flaw they keep secret is a loaded weapon. You are being told this story so that you feel safer. You should feel the opposite. Go look at the leaked founding documents of the organizations that fund OpenAI. Follow the money. The answer is already in front of you.