OpenAI released GPT-6 Cyber on November 5 and the model immediately got to work. Within days it had surfaced 81 previously unknown high-severity vulnerabilities in Firefox and built working exploits for 948 Linux kernel privilege escalation bugs — numbers no human security team could match in the same window. Mozilla has already shipped patches in Firefox 159. Linux maintainers have addressed the reported flaws. If you have not updated, you are running code with known, exploitable holes.
What GPT-6 Cyber Is
GPT-6 Cyber is a specialized version of GPT-6 Astra trained for three tasks: finding vulnerabilities, developing working exploits, and proposing fixes. It is not a general-purpose model — it exists to do offensive security work under controlled conditions. OpenAI rates it Critical under its own Preparedness Framework, the first specialized Cyber variant to earn that designation. Its predecessor, GPT-5.6 Cyber, was only rated “High.” That one-tier difference reflects a meaningful capability jump: GPT-6 Cyber scored 48.5% on ExploitGym versus 32.1% for the previous model — a 51% improvement in the ability to turn a known vulnerability into a working attack.
The Findings: Firefox, Linux, and Brotli
The Firefox disclosure is the most immediately actionable. GPT-6 Cyber identified 81 high-severity flaws, all previously unknown, all patched in Mozilla’s advisory MFSA 2026-118. That advisory covers Firefox 159.0. Mozilla’s Brian Grinstead put it plainly: “GPT-6 Cyber uncovered previously unknown high-severity vulnerabilities…and helped us ship fixes to millions of users.” Thirty-four of the CVEs in that advisory credit OpenAI’s research team as the reporter.
The Linux numbers are larger but less immediate for most developers. The model identified 2,308 potential kernel flaws, reproduced 1,724 of them, and developed 948 distinct local privilege escalation exploits — attack chains that let an unprivileged user gain root. All were reported to kernel maintainers and patched before disclosure. If you are running servers on unpatched kernel versions, check the latest security release for your distribution.
The Brotli finding is the one that should concern Node.js and backend developers specifically. GPT-6 Cyber identified a critical remote code execution vulnerability in Brotli, the compression library used by Node.js, Chrome, and Nginx among others. A full exploit chain was developed and validated. The vulnerability is patched — but only if your runtime has applied the update.
Critical Rating: What It Actually Means
OpenAI’s Preparedness Framework has four tiers: Low, Medium, High, Critical. Hitting Critical is not a marketing label — it triggered a rewrite of the framework itself in August 2026 and introduced mandatory gated access and stricter monitoring for any model at that tier. The practical implication is that OpenAI does not believe GPT-6 Cyber should be in general circulation. The safety metrics are strong (99.93% out-of-scope avoidance without any safety prompts), but the capability ceiling is high enough that unrestricted access would be irresponsible.
Who Can Access It Right Now
Access runs through OpenAI’s Daybreak program. Daybreak Blue covers defensive security work — threat hunting, incident response — using mainstream models with reduced refusals. Daybreak Red unlocks offensive capabilities: penetration testing, red-teaming, exploit development. GPT-6 Cyber goes even further: it requires a separate approval on top of Daybreak Red status. Right now it is available to a select subset of approved Red customers. GPT-5.6 Cyber is available to all Red organizations as an interim option.
Pricing reflects the exclusivity: $20 per million input tokens and $100 per million output tokens — double the standard GPT-6 Astra rates. For more details on the access model, see Help Net Security’s breakdown of the Daybreak program tiers.
The Real Problem: AI Is Collapsing the Patch Window
The bigger story here is structural. Microsoft’s June 2026 Patch Tuesday hit a record 206 CVEs, up from a previous record of 175. OpenAI’s own “Patch the Planet” initiative, running since June 2026, deploys frontier models against 19 major open-source projects and has merged dozens of patches. AI-assisted vulnerability discovery is now operating at machine speed — and the traditional 30-day patch SLA was designed for human-speed discovery.
OpenAI said it directly: “It is clear that we have a limited window to use frontier models to secure software.” The implication is that once these capabilities proliferate to malicious actors, the window for defenders closes. The Daybreak program is OpenAI’s bet that defenders-first access can tip the balance — but only if defenders actually move fast enough once patches arrive. CERT-EU has issued formal guidance on adapting to AI-driven vulnerability discovery timelines.
What You Should Do Now
- Update Firefox to 159.0+ — this patches all 81 vulnerabilities in MFSA 2026-118
- Apply Linux kernel security updates — check your distribution’s security advisory feed for the latest kernel patches covering the reported privilege escalation chains
- Audit your Brotli dependency — Node.js, Nginx, and Chrome-based tools all ship Brotli; verify your runtime version includes the RCE patch
- Reconsider your patch SLA — 30 days is no longer safe for Critical-rated CVEs; compress to 7 days or fewer for high-severity findings
- Monitor security advisories more aggressively — AI-assisted discovery means more vulnerabilities surfacing faster; BleepingComputer’s security feed is a reliable starting point
The model is not publicly available and the access process is deliberate. But the vulnerabilities it found are real, they are patched, and the patches are available today. There is no reason to still be running Firefox 158.
