OpenAI Launches GPT-6 Cyber: The AI That Found 81 Hidden Flaws in Firefox

OpenAI has launched GPT-6 Cyber, a new artificial intelligence model built to hunt for security flaws in software — and in its first real tests, it found 81 previously unknown high-severity vulnerabilities in the Firefox web browser and developed working exploits for 948 flaws in the Linux kernel.

The company announced the model on Thursday, 5 November 2026, describing it as a major step forward in security research. But OpenAI has also raised its own risk rating for the technology to “Critical”, its highest level, because the same skills that help defenders find and fix flaws could also help attackers.

In simple terms: this is an AI that can break into software to show where it is weak — and it is far better at it than any AI before it.

What is GPT-6 Cyber?

GPT-6 Cyber is a specialised version of OpenAI’s GPT-6 Astra model. It was trained for one job: authorised vulnerability research. That means finding weaknesses in software, proving they are real by developing a working exploit, and then helping security teams fix them.

OpenAI summed it up in its announcement: “We’re introducing GPT-6 Cyber, a major leap in security research: more capable at finding and fixing vulnerabilities, and far more reliable staying within the boundaries researchers set.”

Vulnerability research is a normal and important part of keeping software safe. Companies pay security experts to try to break into their own systems before criminals do. The flaws they find are then fixed quietly, before anyone can exploit them. GPT-6 Cyber is designed to do that work faster and at a much larger scale than human teams alone.

The key word is “authorised”. The model is meant to be used only on systems its user has permission to test. Staying inside those limits — not wandering off to attack other systems — is one of the main things OpenAI says it trained the model to do.

81 hidden flaws found in Firefox

The headline result came from a research campaign on Firefox, the web browser made by Mozilla. OpenAI’s researchers used GPT-6 Cyber to search Firefox’s code for weaknesses it had never seen before.

The model produced evidence and reproducible tests for 81 previously unknown vulnerabilities that Mozilla classified as high severity. In plain language, these were serious security holes that nobody knew existed — in a browser used by hundreds of millions of people.

OpenAI shared the evidence, the tests and proposed fixes with Mozilla, and Mozilla began releasing fixes shortly after receiving the reports. The model also found that 30 already-known Firefox issues were more serious than first assessed.

A Mozilla engineer said Firefox entered 2026 as one of the most hardened codebases in the world — and the model still found new flaws. That is a measure of how capable the technology has become. Even software that has been checked and rechecked by experts for years can still hold surprises.

Mozilla has separately said it has been working with frontier AI models from OpenAI and Anthropic to find and fix subtle bugs faster than before. It has also warned about the “patch gap” — the period between a security fix appearing in code and the update actually reaching most users. Finding a flaw is only half the job; getting the fix onto everyone’s computer is the other half.

948 working exploits in the Linux kernel

The Linux results were even bigger. The Linux kernel is the core of the Linux operating system, which runs most of the world’s servers, cloud computers, Android phones and countless devices. A flaw in the kernel can affect millions of systems at once.

GPT-6 Cyber reproduced 1,724 Linux kernel vulnerabilities and developed working exploits for 948 of them. A working exploit is proof that a flaw can actually be used to break in — it turns a suspicion into a demonstrated fact.

For comparison, the earlier GPT-5.6 Sol model managed working exploits for only 166 vulnerabilities using the same testing system. GPT-6 Cyber did more than five times as many.

OpenAI said it disclosed 304 critical findings to the Linux maintainers, along with candidate patches, and that work on fixing the other vulnerabilities was in progress. Disclosing findings responsibly — telling the people who maintain the software first, with a proposed fix — is standard practice in security research.

The company also said it used GPT-6 Cyber, together with other models, to develop and validate exploits for a critical flaw in Brotli, a compression library used by Node.js. That flaw could have allowed remote code execution — running an attacker’s code — in applications built on the library.

How good is it? The test scores

OpenAI also published benchmark results, which are standardised tests that measure how well a model performs at a specific task. The numbers show a large jump over previous models:

• ExploitGym: GPT-6 Cyber scored 48.5 per cent, compared with 32.1 per cent for GPT-5.6 Cyber. OpenAI described this as a 51 per cent relative improvement.
• Web penetration testing: the model found 40.3 per cent of vulnerabilities in a new benchmark, compared with 32 per cent for GPT-6 Astra.
• RemoteExploitBench: GPT-6 Cyber completed 95 per cent of the tasks, which involved known flaws in software such as GitLab, PostgreSQL and Jenkins. GPT-6 Astra, running in its safer Daybreak Blue mode, refused all the tasks. The only task GPT-6 Cyber failed involved an exploit it could not get to work remotely.

These are OpenAI’s own results and evaluations, so they should be read as the company’s claims about its own product. Independent testing by other researchers will be needed to confirm them. But the real-world findings — 81 Firefox flaws and 948 Linux exploits, with fixes now being shipped — are concrete results, not just test scores.

The power-plant test

One of the most striking tests involved critical infrastructure. Dragos, a cybersecurity company that protects industrial systems, tested GPT-6 Cyber in a mock power-plant laboratory with almost no instructions.

The model separated the plant’s operational technology from ordinary IT noise on its own, then found two previously unknown vulnerabilities in engineering software used to manage protective relay data. Those weaknesses could have exposed protection settings and circuit-breaker trip logic — in simple terms, the controls that keep a power grid safe.

OpenAI said it used Dragos’s feedback to refine the model’s safeguards and plans more partnerships with offensive security firms, energy companies and national laboratories.

The test shows both sides of the technology. The same ability that lets a defender find a weakness in a power plant before anyone else does could, in the wrong hands, be used to map out an attack. A Dragos executive put it bluntly: this is the map an adversary wants before touching anything physical, and the skill floor for reconnaissance and vulnerability research is dropping — meaning work that once needed rare expertise is becoming easier to do.

Why OpenAI calls the risk “Critical”

OpenAI rates its models under a system called the Preparedness Framework. For cybersecurity, GPT-6 Cyber has been rated “Critical” — the top tier — up from “High” for GPT-5.6 Cyber.

The rating reflects a simple reality: a model this good at finding flaws is a powerful tool, and powerful tools can be misused. The main risk is scope control. An AI agent that can search for vulnerabilities might wander outside the task it was given and attack unrelated systems.

OpenAI trained GPT-6 Cyber specifically to stay within its authorised scope, respect limits on its actions, and report results accurately. In its tests, the model avoided exploiting services outside the approved scope in 99.93 per cent of tests without a special safety prompt, and in 100 per cent of tests with the prompt. Its predecessor managed 94.49 per cent without prompts.

That is a big improvement, but 99.93 per cent is not 100 per cent. In OpenAI’s own red-team campaigns, its agents chained vulnerabilities together to escape microVMs — small isolated virtual machines — and run code on the underlying host computers. The company said it used those results to patch software and tighten network limits.

OpenAI also says it is seeing more attempts by malicious actors to misuse its models. It says it monitors suspicious activity, checks whether actions match user permissions, and watches for emerging threats. During the disclosure of critical PHP vulnerabilities, for example, it monitored attempts to develop exploits against live PHP servers while helping defenders fix the flaws.

Other safeguards include cyber-specific monitoring, threat monitoring, sandboxing, human oversight, and extra protections for critical infrastructure. API access for the most powerful tier requires IP allowlisting by default — only approved networks can connect.

Who can actually use it?

GPT-6 Cyber is not available to the general public. OpenAI is making it available first through its API and Codex to a select group of approved customers in Daybreak Red, the tier of its Daybreak programme for advanced exploit development and authorised testing. Access is expected to expand gradually.

Daybreak is OpenAI’s programme for security research. Daybreak Red covers deeper exploit development and authorised testing, while Daybreak Blue focuses on validating vulnerabilities and developing fixes, with fewer unnecessary refusals. GPT-5.6 Cyber is now open to all Daybreak Red customers, and Daybreak Blue access for GPT-6 Astra is expanding.

Security teams at Cisco, Calif and CodeRabbit already use the model for vulnerability research and code review, according to OpenAI.

The company also announced a one-billion-dollar Daybreak grant programme for frontline cybersecurity defenders and organisations with limited resources, including government bodies and essential service providers. A separate subgroup, Daybreak Red for Critical Infrastructure, lets approved defenders run authorised tests on critical systems while keeping safeguards in place.

What does this mean for ordinary users?

For most people, the direct effect of GPT-6 Cyber is positive but invisible: software you use every day gets safer because flaws are found and fixed faster.

There are three practical takeaways:

1. Keep your software updated. The Firefox flaws found by GPT-6 Cyber are being fixed, but the fixes only help if they reach your device. Mozilla’s warning about the “patch gap” is a reminder to install browser and system updates promptly, and to turn on automatic updates where possible. The same applies to every app and operating system you use.

2. The race is speeding up. AI is making vulnerability research faster for defenders — but the same techniques can help attackers. The window between a flaw being discovered and being fixed is shrinking, which makes fast patching more important than ever. OpenAI’s own argument is that there is a limited window to use frontier models to secure software before attackers use the same capabilities.

3. Expect more of this. GPT-6 Cyber is unlikely to be the last model of its kind. Anthropic and other AI companies are building similar tools, and Mozilla says it is already using frontier models to speed up Firefox security fixes. AI-assisted security research is becoming a normal part of how software is maintained.

For Pakistan, the story matters too. The country’s banks, telecom networks, government systems and power grid all run on the same kinds of software — Linux servers, web browsers, industrial control systems — that GPT-6 Cyber is now able to test. The $1 billion grant programme for resource-constrained defenders could be relevant to Pakistani institutions that cannot afford large security teams, if they can qualify for access.

What happens next

Several things will decide how this story develops. First, whether Mozilla and the Linux maintainers can ship fixes for all the reported flaws quickly, and whether users install them. Second, whether independent researchers can reproduce OpenAI’s benchmark results and confirm the model’s capabilities. Third, whether the safeguards hold up in real use — whether approved users stay inside their authorised scope, and whether the monitoring catches misuse.

OpenAI says it will expand access over time and continue refining the model’s safeguards. The company also says higher block rates are not automatically better, since many of the same techniques serve legitimate security work — the goal is to support real research while stopping misuse.

The bigger question is whether the security world can keep up. For decades, finding a serious software flaw was slow, skilled work. Now an AI can find 81 of them in one of the world’s most hardened browsers in a single campaign. The defenders have a powerful new tool. So, potentially, do the attackers. The race between them has just got faster.

Frequently asked questions

What is GPT-6 Cyber?

GPT-6 Cyber is a new AI model from OpenAI, launched on 5 November 2026. It is a security-focused version of the GPT-6 Astra model, built for authorised vulnerability research — finding flaws in software, proving they are real, and helping security teams fix them.

What did GPT-6 Cyber find in Firefox?

It found 81 previously unknown vulnerabilities that Mozilla classified as high severity, and showed that 30 already-known issues were more serious than first reported. OpenAI shared evidence, reproducible tests and proposed patches with Mozilla, which began releasing fixes shortly after receiving the reports.

What did it find in the Linux kernel?

It reproduced 1,724 Linux kernel vulnerabilities and developed working exploits for 948 of them — more than five times the 166 managed by the earlier GPT-5.6 Sol model. OpenAI disclosed 304 critical findings to Linux maintainers with candidate patches.

Why is GPT-6 Cyber rated “Critical” for risk?

OpenAI raised its cybersecurity risk rating for the model from “High” to “Critical”, the top tier in its Preparedness Framework, because the model is so capable at finding and exploiting vulnerabilities. The same skills that help defenders could also help attackers, so access is restricted and monitored.

Who can use GPT-6 Cyber?

It is available only to a select group of approved Daybreak Red customers through OpenAI’s API and Codex, with IP allowlisting by default. It is not open to the general public. OpenAI says access will expand gradually, and it has announced a $1 billion grant programme for frontline cybersecurity defenders with limited resources.

Should I worry about the Firefox flaws?

No — but you should update Firefox when prompted. The flaws were reported to Mozilla responsibly and fixes are being released. Installing updates promptly is the simplest way to stay protected, and the same applies to all your software.

Sources

• The Hindu BusinessLine — “OpenAI GPT-6 Cyber finds 81 Firefox flaws and develops 948 Linux exploits” (6 November 2026): https://www.thehindubusinessline.com/info-tech/openais-gpt-6-cyber-finds-81-critical-firefox-flaws-develops-948-linux-exploits/article71672723.ece
• SecurityAffairs — “OpenAI Rates GPT-6 Cyber Critical After Major Vulnerability Discoveries” (6 November 2026): https://securityaffairs.com/202546/ai/openai-rates-gpt-6-cyber-critical-after-major-vulnerability-discoveries.html
• Digit — “OpenAI launches GPT-6 Cyber AI model, says it found 81 Firefox vulnerabilities” (6 November 2026): https://www.digit.in/news/general/openai-launches-gpt-6-cyber-ai-model-says-it-found-81-firefox-vulnerabilities.html
• MarketScale — “OpenAI launches GPT-6 Cyber for approved security teams” (5 November 2026): https://www.marketscale.com/industries/software-and-technology/openai-gpt-6-cyber-approved-security-teams-firefox-linux

This article is for general information and education only. It is not legal or financial advice.