OpenAI released GPT-6 Cyber on November 5, classifying it “Critical” under its own Preparedness Framework — the highest risk tier it has ever assigned to a model. In research testing, the system built working privilege escalation exploits for 948 Linux kernel vulnerabilities and found 81 previously unknown high-severity holes in Firefox. The company’s own “Patch the Planet” dashboard shows less than 25% of those findings have been remediated. The exploits exist. The patches don’t.
What It Found
OpenAI’s researchers turned GPT-6 Cyber loose on the Linux kernel and Firefox. On Linux, the model reproduced 1,724 known vulnerabilities and developed working exploits for 948 of them — 588 of which were previously undiscovered flaws. That’s five times more functional exploits than GPT-5.6 Cyber achieved on the same target. On Firefox, it surfaced 81 unknown high-severity vulnerabilities and determined that 30 previously documented bugs were more serious than their existing severity ratings suggested. Mozilla has started shipping patches; Firefox 159.0 addresses some of them.
The model also identified 10 novel exploitation techniques that security researchers had not previously documented. Those techniques create new detection-evasion pathways — meaning defenders who’ve hardened against known attack patterns may still be exposed.
OpenAI disclosed 304 critical Linux kernel findings with proposed patches. The kernel maintainers have them. The timeline from “disclosed” to “shipped to your server” is the problem.
The Access Asymmetry
GPT-6 Cyber is locked to vetted Daybreak Red customers — approved security firms like Mozilla, Palo Alto Networks, and Cisco, authenticated via IP allowlist, paying $20 per million input tokens and $100 per million output tokens. That’s a reasonable gate for a model rated Critical.
The gate doesn’t stop the threat, though. GLM 5.3, an open-weight model, has no such restrictions. Epoch AI researchers showed it can read a published security patch, reverse-engineer the underlying vulnerability, and build a working exploit for approximately $1,000 per Firefox version. It successfully compromised six recent Firefox releases this way. One of the resulting exploits was a drive-by webpage that exfiltrated files from visiting users without additional interaction.
The aligned model stays in scope. The unaligned model costs less than a used laptop. That is the actual threat model.
The Patch Gap Is Now Your Problem
The patch gap isn’t new, but AI has made it dangerous at scale. The window runs from the moment a security fix appears in a public repository to the moment it reaches users on production systems. It has two phases: code published before a release ships, and users who don’t update after one does.
AI doesn’t need to find a vulnerability from scratch. It can read a published diff and work backwards. Max Justicz, a startup CTO who had early access, found a severe Ubuntu privilege escalation vulnerability in two hours using GPT-6 Cyber. His assessment: “I can only imagine what one could do with more time and access.”
Your deployment pipeline speed is now a security variable. The longer your patch-to-production cycle, the wider the window.
What to Do
- Patch Linux kernels and Firefox now. The disclosed vulnerabilities are real, documented, and accompanied by proposed fixes. The 75% that remain unpatched are the attack surface.
- Compress your patch cycle. Prioritize security updates to production. Automate where possible. The gap between “fix committed” and “deployed” is what AI exploits.
- Apply for Daybreak Blue. OpenAI’s defensive tier gives open-source maintainers and defenders access to AI-powered code review and vulnerability validation. If you maintain infrastructure, get on the list.
- Audit your privilege escalation chains. GPT-6 Cyber excels at chaining small bugs into full privilege escalation paths. Review your code for stacked vulnerabilities — multiple individually low-severity issues that combine into something worse.
The Bigger Picture
OpenAI spent roughly $1.2 million on the Linux kernel research — a rounding error at its scale, and well below what any serious nation-state security operation would spend. The AI cybersecurity market is projected to reach $51.3 billion in 2026, nearly double last year’s figure. The tools are getting cheaper and more capable simultaneously.
The 25% remediation rate on GPT-6 Cyber’s discoveries suggests defenders are already behind. The aligned, restricted model found the exploits and proposed the patches. The unaligned open models are finding their own. The difference is whether the researcher using them has agreed to terms of service.
Read the full OpenAI announcement for Daybreak program eligibility and access details.
