OpenAI GPT-6 Cyber Found 81 Firefox Flaws and Built 948 Linux Exploits

OpenAI has released GPT-6 Cyber, a specialised artificial intelligence model built to hunt for software security flaws, and its first public results are striking: 81 previously unknown vulnerabilities in Firefox and 948 working exploits for the Linux kernel.

Announced on Thursday, November 5, 2026, the model finds weaknesses, builds working attacks against them and proposes fixes. OpenAI rates its capabilities as "Critical" under its safety framework and is releasing it only to approved customers.

Key Points

• OpenAI released GPT-6 Cyber on November 5, 2026, a model built to find software flaws, develop working exploits and propose patches.
• In Firefox, it found 81 previously unknown vulnerabilities Mozilla classified as high severity, and flagged 30 known issues as more severe than thought.
• In Linux kernel testing, it reproduced 1,724 vulnerabilities and built working privilege-escalation exploits for 948 of them, over five times the 166 from GPT-5.6.
• OpenAI disclosed 304 critical findings to Linux maintainers with candidate patches; the testing run cost about $1.2 million in computing.
• It scored 48.5 percent on the ExploitGym benchmark, up from 32.1 percent for GPT-5.6 Cyber, and access is limited to selected Daybreak Red customers.

What Happened

GPT-6 Cyber is a purpose-built security model. Unlike general chatbots, it probes software like a skilled human researcher: reading source code, tracing how data flows through a program, crafting inputs that trigger crashes, and turning those crashes into working exploits that prove a flaw can be used by an attacker.

In the first campaign, it was set loose on Mozilla Firefox. It found 81 previously unknown vulnerabilities, all classified by Mozilla as high severity. For each one, the model supplied evidence, a reproducible test and a proposed patch, and Mozilla has begun shipping fixes. It also re-examined 30 known Firefox issues and judged them more severe than previously reported.

The second campaign targeted the Linux kernel, the core of the operating system behind most servers, cloud infrastructure and Android phones. GPT-6 Cyber reproduced 1,724 known vulnerabilities and built working local-privilege-escalation exploits for 948 of them. Such an exploit lets an attacker with limited access gain full control of a machine. OpenAI said the run cost about $1.2 million in computing, roughly $1,279 per working exploit.

Background and Context

Researchers have long used automated tools to find software flaws, from fuzzers feeding random inputs into programs to analysers scanning code for known patterns. Large AI models change the picture because they can reason about unfamiliar code, form hypotheses about where a flaw hides, and adapt when an attempt fails. That makes them far more effective than earlier tools, but also more capable in the wrong hands.

The leap from GPT-5.6 Cyber is stark. In the same Linux setup, the earlier model produced 166 working exploits; GPT-6 Cyber produced 948. On ExploitGym, a benchmark for exploiting real software flaws, it scored 48.5 percent, up from 32.1 percent. On a new web penetration-testing benchmark, it reached 40.3 percent recall, against 32 percent for GPT-6 Astra, OpenAI's general model.

GPT-6 Cyber: What OpenAI's Security Model Can Do

The model does three jobs security teams normally do by hand. The first is discovery: scanning large codebases for flaws nobody has noticed. The second is exploitation: proving a flaw is real by building a working attack. The third is repair: drafting a patch that closes the hole without breaking the software.

That combination matters because finding a flaw is only half the job. A report that cannot be reproduced is easy for a busy team to ignore. By supplying a reproducible test and a proposed fix, as it did for Mozilla, GPT-6 Cyber turns a suspicion into something a team can act on at once. Mozilla has already begun shipping fixes.

The same capability that helps defenders also helps attackers. A model that can find a flaw, build an exploit and draft a patch can also find a flaw and build an exploit without telling anyone. That dual-use nature is why OpenAI is not releasing GPT-6 Cyber openly, and why access comes with identity checks, organisational security requirements and API IP allowlisting.

Reactions and Quotes

OpenAI frames GPT-6 Cyber as a defensive tool, built to help the people who protect software rather than those who attack it. The company stressed that the Firefox findings were shared with Mozilla with evidence, tests and patches, and that the Linux findings went to maintainers the same way, with 304 critical issues disclosed.

Mozilla began shipping fixes for the flaws, treating them as genuine and serious. The wider security community is likely to split: defenders will welcome a tool that finds flaws faster than human teams, while critics will ask whether a model this capable should exist at all, and whether OpenAI's controls are enough.

Why It Matters

Firefox and Linux are not obscure software. Firefox is used by hundreds of millions of people, and the Linux kernel runs most of the world's servers, the majority of cloud computing and Android on billions of phones. Eighty-one high-severity flaws in Firefox, found by one model in one campaign, show that even well-maintained software can carry serious unknown weaknesses.

The Linux numbers are more striking still. Nearly a thousand working privilege-escalation exploits, produced for about $1,279 each, suggest the cost of finding exploitable flaws is falling fast. For defenders, that is good news: flaws can be found and fixed before attackers use them. But the window between discovery and patching is the most dangerous period in software security, and it may now be closing faster than patching can keep up.

What Happens Next

The immediate step is patching. Mozilla has begun shipping fixes for the Firefox flaws, and Linux maintainers are working through the 304 critical findings OpenAI disclosed. Users should keep browsers and systems updated, as fixes will arrive through normal update channels.

Longer term, the question is how widely GPT-6 Cyber and models like it will be deployed. OpenAI offers it only through its API and Codex, its coding tool, to approved customers who pass identity checks and meet security requirements. If that list grows, it will shape how fast the industry adopts AI-driven vulnerability research. For more on how AI companies build their security businesses, see our report on Anthropic's $17 billion third-quarter revenue and AMD's $10.7 billion charge as OpenAI and Meta chip warrants begin to bite.

Frequently Asked Questions

What is GPT-6 Cyber?
GPT-6 Cyber is a specialised AI model from OpenAI, released on November 5, 2026, built to find software security flaws, develop working exploits and propose patches.

How many Firefox flaws did it find?
It found 81 previously unknown vulnerabilities in Firefox, all classified as high severity by Mozilla, and flagged 30 known issues as more severe than previously reported.

What did it do with the Linux kernel?
It reproduced 1,724 vulnerabilities and built working local-privilege-escalation exploits for 948 of them, more than five times the 166 exploits managed by GPT-5.6.

How much did the Linux testing cost?
About $1.2 million in computing, roughly $1,279 per working exploit, according to OpenAI.

Can anyone use GPT-6 Cyber?
No. OpenAI is releasing it only through its API and Codex to selected approved Daybreak Red customers, with identity checks, security requirements and API IP allowlisting.