OpenAI launches GPT-6 Cyber: a model that found 81 unknown Firefox flaws and built 948 working Linux exploits
OpenAI has released GPT-6 Cyber, a model built to hunt for software vulnerabilities and turn them into working exploits — and its first field tests read like a stress test for the entire security industry. In Firefox it surfaced 81 previously unknown flaws Mozilla rated high severity. Against the Linux kernel it reproduced 1,724 vulnerabilities and built working exploits for 948 of them. The catch: the model carries OpenAI's top Critical cyber rating, and for now it is available only to a small circle of vetted security teams.
- Firefox: 81 previously unknown vulnerabilities, all rated high severity by Mozilla; Mozilla has shipped fixes for many of them. The model also showed 30 previously reported issues were more serious than first assessed.
- Linux kernel: 1,724 vulnerabilities reproduced and working local-privilege-escalation exploits built for 948 of them, 588 previously unknown; 304 critical findings disclosed to maintainers with candidate patches.
- Benchmarks: 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra. Without deployment safety instructions it went out of scope in 0.07% of graded runs, versus 5.51% for GPT-5.6 Cyber.
- Rating: Critical — the top cybersecurity tier of OpenAI's Preparedness Framework. GPT-5.6 Cyber was rated High; GPT-6 Astra shares the Critical rating.
- Access: API and Codex only, for a select group of approved Daybreak Red customers, with identity verification and API IP allowlisting. OpenAI says access will expand to all Daybreak Red customers over the coming weeks.
The launch, announced November 5, turns a September report into a shipping product. We covered GPT-6 Cyber in September as a rumored DevDay preview; OpenAI has now confirmed it with a full announcement, a system card, and a set of results that are hard to dismiss. GPT-6 Cyber is a cyber-specialized variant of GPT-6 Astra, tuned for authorized security work: finding vulnerabilities, proving they can be exploited, and helping teams fix them. OpenAI says the model is trained to stay inside the boundaries researchers set — a claim the company is now backing with numbers rather than adjectives.
The headline numbers are the ones that will travel furthest, and they deserve a careful read. These are OpenAI's own evaluations of its own model, not independent verification. But the underlying work is concrete: real flaws in real software, disclosed to the teams that maintain it, with patches already landing in at least one case.
What the model actually found#
The Firefox campaign is the cleanest result. OpenAI's researchers used the model to produce reproducible evidence for 81 previously unknown vulnerabilities that Mozilla classified as high severity, along with proposed patches — and Mozilla has shipped fixes for many of them. The model also re-examined 30 previously reported issues and showed they were worse than first assessed. That second detail matters: finding new bugs is one skill; correctly judging the severity of known ones is another, and it is the one that decides which flaws get fixed first.
The Linux kernel work ran at a different scale. Using an automated research pipeline against the latest kernel development code, the model reproduced 1,724 vulnerabilities and developed working exploits for 948 distinct local-privilege-escalation flaws — attacks that let an unprivileged local user obtain root access in the test environment. Of those, 588 were previously unknown. OpenAI says it disclosed 304 critical findings to kernel maintainers together with candidate patches, and withheld the technical details of its techniques. The figures do not mean 948 new remotely exploitable flaws have been released into the wild; they mean the model can take a candidate flaw, prove it, and weaponize it inside a test environment at a pace no human team matches.
| Target | What GPT-6 Cyber did |
|---|---|
| Firefox | 81 previously unknown high-severity vulnerabilities found, with reproducible evidence and proposed patches; 30 known issues re-rated as more serious |
| Linux kernel | 1,724 vulnerabilities reproduced; working exploits built for 948 local-privilege-escalation flaws, 588 of them previously unknown; 304 critical findings disclosed with candidate patches |
| ExploitGym | 48.5% — up from 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra |
A sharper tool that stays in bounds#
Capability is only half the announcement. The other half is alignment: a model that can build exploits is only useful to defenders if it stays inside the scope it was given. On ExploitGym, a benchmark that tests whether a model can exploit a specified vulnerability in a remote target, GPT-6 Cyber scored 48.5%, against 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra. And in OpenAI's own scope tests, the model went out of bounds in 0.07% of graded runs without deployment safety instructions, compared with 5.51% for GPT-5.6 Cyber — a roughly eighty-fold improvement in staying where it was told to stay.
That combination — more capable and more obedient — is what earned the Critical rating. Critical is the top cybersecurity tier in OpenAI's Preparedness Framework, reserved for models judged able to find and exploit previously unknown vulnerabilities in hardened systems without human direction. GPT-5.6 Cyber was rated High. GPT-6 Astra, the model GPT-6 Cyber is built from, was the first to reach Critical when it shipped in September; GPT-6 Cyber now shares the rating. OpenAI says it has added safeguards for industrial controls, essential services, election systems, and other high-risk targets.
The gate stays up#
Access is where the announcement gets restrictive. GPT-6 Cyber is available through the API and Codex only to a select group of approved customers in Daybreak Red — the tier of OpenAI's Daybreak program for offensive-style work such as penetration testing, exploit development, and red teaming. Customers go through identity checks and organizational security requirements, and API requests must come from approved networks by default. OpenAI says access will expand to all Daybreak Red customers over the coming weeks, and it is opening the previous GPT-5.6 Cyber to all Daybreak Red customers at the same time.
That is a deliberate choice, and it is the same one every frontier lab has converged on this year: the sharpest cyber capabilities ship through gates, not storefronts. OpenAI says security teams at Mozilla, Palo Alto Networks, Calif, and Cisco are already putting the model to work, alongside its own researchers. The company is also expanding the Daybreak Blue alpha — the defensive tier, for validating vulnerabilities and fixes — for GPT-6 Astra to more organizations and individual researchers.
What to watch#
The question the launch raises is not whether AI can find vulnerabilities faster than humans — that is now settled — but whether the rest of the system can keep up. Discovery is accelerating; patching is not. According to one report, OpenAI's own Patch the Planet dashboard shows fewer than a quarter of reported vulnerabilities have been fixed. A model that surfaces flaws at machine speed creates a backlog problem for maintainers, and the backlog is where the risk actually lives: a known, unpatched flaw is more dangerous than an unknown one.
Watch three things. First, whether Daybreak Red access widens on the promised timeline — and who gets in when it does. Second, whether Mozilla and the kernel maintainers clear the disclosed findings, and how long the fixes take to reach users. Third, how regulators respond to a second Critical-rated model in as many months, and whether other labs answer with their own gated cyber variants. The September preview is now a product. The next question is who gets to use it — and how fast the industry it is meant to defend can absorb what it finds.
Sources#
- OpenAI — “Introducing GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents” (November 5, 2026)
- MarketScale — “OpenAI launches GPT-6 Cyber for approved security teams” (November 5, 2026)
- The Hindu BusinessLine — “OpenAI GPT-6 Cyber finds 81 Firefox flaws and develops 948 Linux exploits” (November 6, 2026)
- Technology.org — “OpenAI Launches GPT-6 Cyber With Critical Rating” (November 6, 2026)
- CIO Africa — “OpenAI's New Cyber Model Found 81 Firefox Flaws” (November 6, 2026)
- MLQ — “OpenAI Releases GPT-6 Cyber After Model Finds 948 Linux Kernel Exploits” (November 5, 2026)