Mozilla's GPT-6 Cyber Agents Found and Fixed 8x More Firefox Bugs in a Year
Mozilla has connected AI agents running on GPT-6 Cyber, OpenAI's cybersecurity model, to its Firefox security work. From 2025 to 2026, the number of bugs found and fixed rose roughly eightfold. The agents decide which finding to examine first, verify that a vulnerability is real, and search for the same mistake elsewhere in the codebase. Some known bugs also turned out to be more dangerous than expected. The takeaway: AI is useful for reordering the task queue, not just for finding holes.
AI-processed from YouTube: OpenAI; edited by Hamidun News
Mozilla says that after connecting AI agents built on GPT-6 Cyber (OpenAI's model for cybersecurity tasks), the number of security bugs found and fixed in Firefox rose roughly eightfold between 2025 and 2026. This covers the whole process, from spotting suspicious code to shipping a fix. Picture a building maintenance service that used to inspect apartments by hand, and now has an assistant: it points out where the leaks are worst, checks whether a leak is real, and immediately checks the same pipes in neighboring buildings. That is roughly how this works inside browser code.
A browser is one of the most attacked programs on any computer. Firefox consists of millions of lines of code written over decades, and it can contain flaws that let an attacker run foreign code or steal data. There are always fewer security specialists than bugs. That is why Mozilla's account matters to anyone responsible for security in a large project.
In an OpenAI video, Clint Gibler of OpenAI talks with Brian Grinstead of Mozilla. The headline figure from the video description: across the whole security process, the number of bugs found and fixed grew roughly eightfold between 2025 and 2026.
What the agents actually do
According to the description, AI is built into more than one stage. It is embedded across the entire vulnerability lifecycle. The first step is prioritizing investigations. There are many signals of possible problems, and the team must decide which to start with. It works like sorting a hospital triage queue: the most serious cases are seen first.
The second step is testing and verification. A suspicion is not yet a vulnerability: the team has to confirm that the error reproduces and can actually be exploited. This is where AI tools often produce false alarms, so without verification the stream of findings turns into noise.
The third step is searching for related bugs. If code is unsafe in one place, programmers have likely used the same pattern elsewhere. The agent looks for similar patterns across the whole codebase, so instead of a single fix, the team closes an entire class of problems at once.
The fourth step is helping engineers with fixes. Decisions about patches stay with people, but AI speeds up preparing them.
Old bugs turned out to be more dangerous
A separate part of the conversation concerns bugs that were already known. GPT-6 Cyber showed that some of them may be more serious than the team first assessed. This helps decide which issues to fix first. The practical lesson: AI is useful not only as a hunter of new holes, but also as a second opinion on what has long sat in the tracker. A bug that sat for years at low priority may be the one that needs fixing today.
What to keep in mind
The source is the description of an OpenAI video, not an independent report. It lacks the details needed to evaluate the result: how "bugs" were counted, how many of the findings were critical, what share were false positives, and how much human time went into verification. An eightfold increase speaks to volume, not to how much safer the software has become. There is also no information on the cost or access terms for GPT-6 Cyber.
In addition, Mozilla is a special case: it has decades of accumulated security culture. The description emphasizes that the AI agents are plugged into exactly that experience. A team without established processes is unlikely to get the same effect just by running a model against its repository.
Takeaways for teams
If you are responsible for product security, this story offers three practical ideas. First, connect AI not to a single task but to a chain: prioritization, verification, searching for similar bugs, and preparing fixes. Second, keep a verification stage so you do not drown in false positives. Third, periodically run old "known" bugs through the model: reassessing priorities may deliver more value than hunting for new ones.
Brian Grinstead goes into more detail about Mozilla's plans for protecting the browser in the video itself. For now, this is more a declaration of a new standard for handling vulnerabilities than a ready, independently verified recipe.
- Nov 21, 2026llama.cpp 1.0.0: Stable API for Developers and 35 Old Functions Removed
- Nov 21, 2026OpenAI Launches ChatGPT Wallet: Agents Pay by Card, Approved With Face ID
- Nov 21, 2026Krea Opens Beta of Krea Bots: 9 AI Bots Produce a 15-Second Commercial
- Nov 21, 2026ChatGPT Wallet Lets OpenAI's Dots Agents Pay by Card, But Every Purchase Requires Face ID
Key Questions Answered
What the agents actually do?
According to the description, AI is built into more than one stage. It is embedded across the entire vulnerability lifecycle. The first step is prioritizing investigations. There are many signals of possible problems, and the team must decide which to start with. It works like sorting a hospital triage queue: the most serious cases are seen first. The second step is testing and verification. A suspicion is not yet a vulnerability: the team has to confirm that the error reproduces and can actually be exploited.
What are the key details on Old bugs turned out to be more dangerous?
A separate part of the conversation concerns bugs that were already known. GPT-6 Cyber showed that some of them may be more serious than the team first assessed. This helps decide which issues to fix first. The practical lesson: AI is useful not only as a hunter of new holes, but also as a second opinion on what has long sat in the tracker. A bug that sat for years at low priority may be the one that needs fixing today.
What to keep in mind?
The source is the description of an OpenAI video, not an independent report. It lacks the details needed to evaluate the result: how "bugs" were counted, how many of the findings were critical, what share were false positives, and how much human time went into verification. An eightfold increase speaks to volume, not to how much safer the software has become. There is also no information on the cost or access terms for GPT-6 Cyber.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.
