OpenAI's GPT-6 Cyber is the best bug hunter yet, and you can't have it
OpenAI's new security model found 81 high-severity Firefox bugs and built working exploits for 948 Linux kernel flaws. It's restricted to a vetted tier of its Daybreak program, and that's on purpose.
On Thursday OpenAI announced GPT-6 Cyber, a model tuned to find and fix exploitable flaws. The numbers are hard to ignore. BleepingComputer reports it found 81 previously unknown high-severity vulnerabilities in Firefox, and Mozilla has already shipped many of the patches. In the Linux kernel it flagged 2,308 potential flaws, reproduced 1,724 and wrote local privilege-escalation exploits for 948 distinct bugs. On the ExploitGym benchmark it exploited 48.5% of vulnerabilities. Its predecessor, GPT-5.6 Cyber, managed 32.1%.
OpenAI rates the model "Critical" for cybersecurity capability under its Preparedness Framework, one step above the "High" it gave the previous model. So it isn't public. Only selected customers in Daybreak Red, the offensive-research tier, get access for now, and OpenAI says it plans to open it to all Daybreak Red customers soon. Verified defenders on Daybreak Blue get fewer restrictions for investigating and validating fixes. In OpenAI's words, there is "a limited window to use frontier models to secure software."
TechCrunch picked up the tension in that framing. A researcher who used Daybreak Blue to find a severe Ubuntu server flaw says OpenAI recently "nerfed" the tier. Defenders also argue that gated access leaves them behind attackers, who can run open-weight models with the guardrails stripped. Costs matter too. TechCrunch puts the Linux kernel sweep at about $1.2 million in compute, and that researcher spent roughly $5,000 finding several exploits.
For a mid-size Canadian shop, the practical effect is that patch volume is about to rise. When one model can drop nearly a thousand kernel fixes into the pipeline, the open-source components under your stack will churn faster, and your patch backlog will grow with it. The same capability will also show up on the attacking side, as this week's ARTEX story shows.
Sources (4)
- BleepingComputer OpenAI's new GPT-6 Cyber AI finds 81 Firefox flaws, 948 Linux bugs — Nov 6, 2026 ↗
- TechCrunch OpenAI released a new cybersecurity model. Is that enough? — Nov 5, 2026 ↗
- OpenAI GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents — Nov 5, 2026 ↗
- OpenAI A thousand fewer paths to root — Nov 4, 2026 ↗