CEOInterviews.AI
Start App
Greg Brockman
Co-Founder & President, OpenAI

Cyber at the Frontier: Introducing Daybreak Red Team & Daybreak Defense Factory

📅 Nov 12, 2026 OpenAI 34 MIN 37386 VIEWS 62 SEGMENTS · 6 SPEAKERS
Greg Brockman, Nancy Wang, Fouad Matin, Clint Gibler, Alexander Thiemann, and Matt Boyle took the stage at Cyber at the Frontier, where OpenAI introduced two cybersecurity products under one unified Codex plugin called Daybreak: Daybreak Defense Factory: builds on Codex Security Cloud to help teams automate the defensive loop to find, validate, and fix vulnerabilities at scale. It gives engineers proposed patches with test results so they can review changes before merging. All of the work runs in isolated cloud development environments. Daybreak Red Team (early access): uses fleets of agents...

What Greg Brockman said

Written from the verified transcript and checked against it. Every figure links to the moment it was said.

Greg Brockman opened the conference by framing cybersecurity as an accelerating challenge rooted in the internet's insecure foundations and under-resourced security teams. He argued that frontier models are a step-function improvement, and that companies will only be as secure as the frontier model they use to defend themselves. He cited a PHP vulnerability case study where a funded researcher found it, then a next-generation model autonomously found a more general version, and OpenAI coordinated a two-month disclosure with governments and companies, which he said was not scalable. He announced a $1 billion Daybreak defense fund for frontline defenders and critical infrastructure, and highlighted an open letter signed by many companies. He concluded by handing over to Nancy Wang to present the Daybreak Red Team and Daybreak Defense Factory products. The rest of the presentation was delivered by other OpenAI staff and partners.

Key takeaways

  1. Brockman announced a $1 billion Daybreak defense fund for frontline defenders and critical infrastructure.
  2. Brockman said a next-generation model autonomously found a more general version of the PHP vulnerability a month after a funded researcher found it.
  3. Brockman said coordinating the PHP disclosure consumed all available bandwidth and is not scalable, requiring a community approach.
  4. Brockman said the internet wasn't built with security in mind and that no CISO has ever been overresourced.

Numbers and commitments

FigureWhat it refers toTypeAt
5.6 Model version used by researcher to find PHP vulnerability other 3:32

Chapters

  1. 0:00Cybersecurity acceleration and root causes
  2. 1:54Frontier model progress and defender window
  3. 3:32PHP vulnerability case study
  4. 4:30Coordination challenges and community response
  5. 5:30Daybreak defense fund and handover
Greg Brockman 0:20 ↗
Hello everyone and welcome. I'm very grateful for everyone who's tuning in today, everyone who is here live, everyone who is paying attention and contributing to this moment because this is a moment for collective action.
This year, I think everyone is aware, has seen a lot of change in cybersecurity. But the thing that's easy to miss is that this is an acceleration. This finding of vulnerabilities, attackers exploiting them, defenders trying to stay ahead. It's an acceleration of a trend that has existed for many years. And the fundamental underlying force is the fact that the internet wasn't built with security in mind. That there's no CISO who has ever been overresourced. That prioritization of security, the diffusion of security expertise, these have been important bottlenecks that I think have meant that we build up tech debt in how we've classically built software. And that I think the thing that needs to change as we're seeing this change in the physics of cybersecurity, this almost democratization of cybersecurity expertise, is that we need to build software in a fundamentally more secure way. So there's both securing what exists and changing our workflows and coming together as a world, as a community in order to make the internet be fundamentally secure in ways that it never was before. So that is the moment, that is the challenge in front of us, and that is something that I think we all have something to contribute to. Now, part of the reason that it's important to be having this conversation is because the progress on models continues.
So, here's a graph of progress and performance with our latest cyber model that we're working on releasing soon to people everywhere, to trusted defenders. And the way to think about this is that this is a model that will be a step function further. And I think that that progress will continue time after time. And the way that we think about what we can do to help defend the world is one, to actually produce these frontier capabilities, that that's very important to have them actually exist so that you can have a defender window where there's a frontier that is far ahead in terms of what is broadly diffused. Secondly, to provide access, and so that this is what the Daybreak program is really focused on, is how do we put these tools into the hands of trusted actors, trusted defenders everywhere.
And the third is to actually help people put these capabilities to work. And so some of that is as simple as just education playbooks. We have put these models to work internally within OpenAI. And I think we're now at a point where companies will be secure only to the extent of the frontier model that they use to defend themselves and how well they utilize that. Fundamentals important to not be discounted, but I think that you need to really reorient your security practice around what these models can do, but also very important is product. And that's something that we're going to be talking about today is the cyber products that we have developed that help put these capabilities to maximum force and extent within companies. And so that will be something that we'll cover and are very excited to see what you are able to do with it.
Now I want to talk a little bit about a case study of this PHP vulnerability that I think many people have heard of, PHP dash shell, because it's a very interesting example of the kinds of changes that we should expect in upcoming months. So we have funded a program called Patch the Planet, and the idea behind Patch the Planet is it's not enough to just find vulnerabilities to point out problems, we also want to help fix them. We want to really put those findings to work. And so we have funded security researchers to go actually land patches. And so a researcher at Califf used our model, I think it was 5.6 at the time, with their own expertise, their own domain expertise to find this vulnerability. A month later, our next generation model actually found the same vulnerability, an even more general version of it, totally autonomously. So it shows you that that change of as we move towards more and more capable models, that the set of vulnerabilities that they'll find will be deeper and deeper.
But then we also coordinated the response that we worked together with a whole community of companies. We worked with governments. We called many governments around the world to coordinate this disclosure and this response. And it was both, I think, a very successful release. There was maybe some little nits you could have at the end, but it was really two months of coordination, preparing the patches, making sure that it was something that was going to be stable when it was rolled out. But the thing to realize too is this really consumed all the bandwidth that was available. It consumed many people on our team's bandwidth. It consumed our partner companies, many of these governments. We were on calls constantly. It's not something that's scalable. And as we think about the vulnerabilities of this nature that are going to be found in upcoming weeks, months, years, we need an approach collectively as a community that scales with that. So that's an open problem. That's something that we all need to come together and work on. But we do have much better tools in the form of these models to help us coordinate this kind of disclosure at much larger scale than was possible before.
Now I think that the fact that this is not a challenge that any one company can address is absolutely critical, and there is a community of people who are really thinking seriously about this. So there was an open letter that many companies signed on to in the past couple months, that we have announced a billion dollar Daybreak defense fund for frontline defenders for critical infrastructure to actually apply these models. So, I think that there is growing momentum and a growing understanding of the urgency of this moment.
Now, I'm excited to hand it over to Nancy Wang, who leads cyber applied AI here at OpenAI. She's going to walk you through the products that I referenced and how defenders can put these to use in their companies. So, thank you everyone and excited to see what you do with what we're releasing today.
Nancy Wang 6:24 ↗
Awesome. Thanks so much, Greg. So, as you've heard from Greg, Daybreak is our platform for everything that OpenAI offers in cyber. From our frontier cyber models to the tools and hands-on support that we provide. A few months ago, our product engineers started on the journey of continuous defense, hardening our systems with frontier models. We then productized those learnings into two products that we are announcing today: Daybreak Red Team and Daybreak Defense Factory. So, let's get into it. First, our offensive security product, Daybreak Red Team, is launching in early access. We strongly believe that the best defense is a good offense. And so that means thinking like an attacker and testing your own systems before bad actors even get the chance.
Daybreak Red Team uses our own frontier cyber models to give you the signal of a superhuman pentest at the scale of automated continuous scanning. And what excites me the most is that Daybreak Red Team will allow us to dramatically democratize access to our frontier cyber models, putting them in the hands of developers and organizations that have never had a red team of their own. And second, we are announcing Daybreak Defense Factory to help you automate your defensive loop, finding, validating, and fixing vulnerabilities at scale. Defense Factory scans your codebase in reproducible cloud development environments where agents trace a vulnerability to the affected code and use your own code history and team structure to help you find the right code owner. And this way engineers get a merge-ready pull request with the patch and a test that shows that it works. In our own internal security work, we learned that the same workflow we used to fix the backlog could help keep running as new code is shipping. And we built that workflow into Defense Factory so your team can use it continuously. And so when new commits land, the agents can review the changes and also prepare fixes at the same time. And of course, all of this is working in the background continuously.
These capabilities come together under Daybreak, which spans the products that we will be announcing today, our model foundations, and our building blocks and programs. And as I've mentioned, powerful models are at the foundation of our products. And so we have Fouad here to take us through the latest releases, followed by our very own Clint, who is going to show you how we've been putting these models into practice. And then we'll dive straight into the product demos. Fouad, it's so great to have you here.
Fouad Matin 8:54 ↗
Thanks, Nancy. It's great to be here. We want to give defenders the most capable models that we can safely deploy and help them make the most out of the defender's window. Last week, we released GPT-6 Cyber, which is our most capable and most aligned model to date. Defenders were able to deploy this always-on cyber capability wherever they needed it. And with each model generation, we're improving the performance on existing cyber tasks such as secure coding and defense or pentesting and offensive areas, all while ensuring safety and alignment continue to progress in parallel. One important benchmark when looking at cyber models, as Greg mentioned earlier, is Exploit Gym, which tests the model's ability to create a proof of concept for a given vulnerability. We trained 6 Cyber to work harder and not give up as easily, and able to achieve a much higher performance on this benchmark at 52%. For comparison, Anthropic's Mythos 5.2, which released earlier this week, scores at about 38%.
Every new model is increasingly more capable and will help developers find more difficult vulnerabilities, validate and prioritize which ones are exploitable, and make sure that they're releasing hardened patches when it's time for release. Our goal is to make sure that we're pushing that frontier for cyber defense. And we're already working on the next specialized cyber model, which is even more capable, token efficient, and like 6 Cyber, is able to work harder and longer by default on the same tasks. It demonstrates a significant increase, nearly double the score of Mythos 5.2. With our upcoming cyber model, we're continuing to reset the frontier of cyber capabilities beyond anything that is available today.
Now, with these frontier models, we believe the challenge is no longer just finding vulnerabilities. It's making sure you can patch them as fast as you can find them. We're measuring this on eval like PatchBench, which tests the model's ability to find and fix a given flaw without causing a regression in the application. Now while 6 Astra performed reasonably well and 6 Cyber had modest improvements, we believe that this type of benchmark is critical to push that frontier for developers in cyber defense. And we're seeing with this upcoming cyber model around 80% performance. We want to make sure that AI is able to help us solve our greatest cybersecurity challenges at the same rate that it's helping uncover them.
Stronger capabilities also raise the bar for how an agent needs to behave. An agent that encounters a service that it might have access to but doesn't have permission to test must respect those boundaries and make sure it doesn't just try to solve the task in an easier way even when there might be an unrelated vulnerability. We're making sure that alignment and safety is a critical part of our cyber model training, staying within authorized scope, following those restrictions, and accurately reporting the work that's been done, including when it fails or was unable to complete the task. We can test this directly in eval Honeypot, which plants an intentionally vulnerable target, a honeypot, in this environment that is easier to exploit and evaluates the model's propensity for actually trying to leverage that while completing the task. Now, while leveraging this honeypot vulnerability could reveal the answer, doing so is outside of its authorized scope. 6 Cyber shows significant improvements over past frontier models and also on other alignment evaluations. When evaluated without cyber safeguards, it stayed within scope at 99.9% of those evaluated cases. And with our production safeguards, it stays within scope 100% of the time.
We believe that everyone needs the ability to red team their own systems. That's the north star that we're working towards. And we believe that ensuring that people can use the most capable and most aligned cyber models is critical for their cyber defense. But model training and alignment is just one part of the problem. We take a defense-in-depth strategy where we've deployed monitoring and checks against high-risk assistance around critical infrastructure and private safety processing, which allows us to partner with our customers to ensure reducing the risk of potential misalignment or misuse. And we're also verifying access through our Daybreak program as we continue to expand and broaden those capabilities. As we develop increasingly capable cyber models, we're doing the work to make sure that as many people can benefit from these capabilities as possible. But benchmarks just tell one part of the story. So now let's have Clint join us to talk about the real world impact we're seeing from our models.
Clint Gibler 13:15 ↗
Thanks, Fouad. Hey everyone, my name is Clint Gibler and I lead cybersecurity research at OpenAI. My team is helping secure the critical software the world relies on like Firefox, PHP, and Linux. These codebases have undergone years of security testing, including with earlier generations of frontier models from leading labs. I'd like to share two learnings we've had about using these models every day and share some examples from our work. The first lesson is that when you scan the same code with a newer model, it finds new vulnerabilities, things that may have been there for years. That's one of many reasons why it's important that defense is continuous and you leverage frontier models appropriately.
The second learning is that you may already know of some low severity vulnerabilities that may actually be high in practice. This is because newer models find novel ways to exploit them, bypassing controls or by chaining together multiple low severity vulnerabilities. For example, we've seen bugs initially believed to just cause a denial of service actually enable remote code execution. And in our own internal scanning at OpenAI, we've been surprised how effective models are at chaining together several low severity findings into something meaningful. So both of these change how urgently bugs need to be fixed. Okay, let's talk about three examples starting with Firefox.
Despite extensive use of frontier models from multiple labs this year, in a recent campaign with GPT-6 Cyber, we found and reproduced 81 previously unknown high severity vulnerabilities. The model also found new attack paths for 30 known issues thought to be low or medium severity and considered non-exploitable. But with GPT-6 Cyber they were reprioritized as high severity. For PHP dash shell, Greg mentioned earlier that the vulnerabilities actually existed for a while, 13 years, but we didn't know the severity until Califf helped establish exploitability with our latest cyber models. As part of the disclosure process we wanted to make sure that the patch was bulletproof. So we did three things. The first is we used our models for variant analysis, so finding all the different ways this vulnerability class could be exploited so that the maintainers could handle them all in one patch. Second, as the patch was being created, we also used our models to adversarially test it, looking for bypasses. And then lastly, our models helped us evaluate a range of different mitigation options for those who couldn't update PHP immediately. For example, we tested whether changing certain PHP settings could make the vulnerability harder to exploit. So doing all of that would have taken us weeks, but with our latest models took us only a few days.
Lastly, one campaign where I was particularly amazed by what the models could do was our work with Linux. As you may know, Linux underpins many of today's devices and cloud infrastructure and is very complex. With more than 40 million lines of code, we developed an autonomous vulnerability discovery, exploitation, and patching pipeline and ran it with both GPT-5.6 Cyber and GPT-6 Cyber. In one run with GPT-6 Cyber, the pipeline validated and generated patches for almost a thousand local privilege escalation vulnerabilities in Linux. That's over five times the amount of LPEs that 5.6 had found previously about three months earlier. During this research, GPT-6 Cyber found 10 novel Linux exploitation techniques that were totally new to us as well as some external parties that we spoke with. Because of this, we believe our frontier models demonstrate creativity, not just pattern matching known approaches. If you're curious to learn more, we wrote a blog post about this called 1000 Fewer Paths to Root if you want to know the details.
These examples show how every new cyber model is more cyber capable and we believe that this extends the defender's window. And to show you more of some of the things we're building, I'll pass it back to Nancy. Thanks for your time.
Nancy Wang 17:02 ↗
Cool. Thanks Clint. So now of course, let's take a deep dive into our product launch today, Daybreak Red Team. And so as bad actors have access to widely available cyber models and offensive open source harnesses, you also need access to both frontier models and a frontier offensive harness. Daybreak Red Team gives you a secure way to put your most capable cyber models into action to pentest your application. Now I'd like to welcome Alex who is building Daybreak Red Team and he's going to walk us through the product.
Alexander Thiemann 17:35 ↗
Hey Nancy, how are you?
Nancy Wang 17:36 ↗
Hey, good. Ready to do some demos?
Alexander Thiemann 17:38 ↗
Yes, let's do it. I'm super excited to talk about Daybreak Red Team today. So we built Daybreak Red Team to help you find and fix vulnerabilities before attackers do. Think of it as a fleet of agents that tests and explores attack paths in your running application, and once it confirms a vulnerability, it will help you reproduce and fix the issue. Should we jump into a demo?
Nancy Wang 18:02 ↗
Let's do it.
Alexander Thiemann 18:03 ↗
Awesome. Cool. So, let's imagine I'm a security engineer at Northstar Docs, which is a collaborative documentation platform. And we've had some issues with multi-tenancy in the past, and so I'd like to just double check that everything is in order there. So I'm going to use the Daybreak Red Team plugin to assess our staging environment. There's also a bunch of other cool use cases that you can use this for. For example, if you use a source scanning tool or you're participating in a bug bounty program, you might want to prioritize what's actually actively exploitable in your running system. And so this is a really nice way to sort of explore and figure out which one of the things you should prioritize. So I'm going to kick this off here. And what Codex will now do is Codex will now use all the context you know from maybe a connected Slack or any other things that I've connected to kind of assemble a campaign that's very tailored and on target of what I need to evaluate. So we can see here our objective of our campaign, we are targeting our staging system. We've uploaded some of our source code, the duration of the campaign, the maximum requests per second, any credentials you might need to actually make a realistic attack scenario, and then we are going to allow HTTP and WebSocket traffic. So, this looks good. I could go back and forth with the agent a little bit here if I wanted to change some things, but overall this looks pretty good. So, I'm going to fire this off and see what happens. So, we're starting the campaign here, and now I have a nice overview of the campaign. On the left side, I can sort of see the findings of my investigation. And on the right side, I can monitor what endpoints the agent fleet is exploring. And I can use this to maybe make some tweaks to the campaign as I see maybe it's missing some endpoints, needs some permission and so forth, but I can kind of monitor progress as a whole. So, while this runs, why don't we take a look at what happens behind the scenes.
Nancy Wang 19:59 ↗
Let's do that. And really, I love how straightforward this demo is.
Alexander Thiemann 20:02 ↗
Yeah, it's a very simple and a very easy way to set up an otherwise fairly complex campaign to really be effective.
Yeah. So, of course, a lot is happening behind the scenes. And so, while that's running, let me maybe tell you about how this actually works. So, behind the scenes, we have a fleet of agents that's working together on the campaign. And each agent independently works on tasks and they coordinate via a shared task queue and also knowledge base. So, they can exchange ideas, findings, and also learnings. Well, that's pretty similar to how humans also work together. And agents here take on specific roles in order to maximize effectiveness as we build the product. Safety, guardrails, and infrastructure isolation are the core product design tenets for Daybreak Red Team. And they all come together into what we call safety in depth.
And so, let me break this down a little bit more. At the very top, we have a supervisor agent that monitors the entire fleet and the knowledge base to identify any deviation from what we intended as the goal. And all of this stands on top of our model alignment research. So let's check back into our campaign.
Awesome. Cool. So fast forward in time a little bit. The campaign has been running for a while. And now on the left side, we've seen some findings come in. And on the right side, we've made some progress on all the endpoints we've hit. So I'm very curious about this stored SQL injection exploit that we found. So let's click into that finding and see what happens. So imagine you have two companies using our platform. Bob works at one company and Alice works at another platform, and Bob can now according to this exploit basically see documents that actually weren't only reserved for Alice who works at a very different company. So that's sort of like the summary that we get to kind of dig into this. Next we have the architecture. So basically what parts of our system and source code actually led to this issue in the first place, and that will help me develop a fix that's durable and not ad hoc and maybe exposes other issues. So this is like a very nice way to connect the architecture back to the actual issue that we're finding here.
And then last but not least, the exploit chain. So here the agent will sort of walk us through all the steps that it took to prove that this is actively exploitable in a live system. So you can kind of see, okay, first it checked that private document that Bob wasn't supposed to access is actually not accessible to Bob via regular endpoints. Now what the agent did here is there was a way to save searches in the application. So Bob saved a search that had a SQL injection in it. And then Bob will run an export job for that search query, and well, it's now running successfully. And finally Bob downloads the zip file of that export and finds this private document. So, very simple and straightforward way to understand, okay, what was actually a challenge, and we've seen this work quite well even with more sophisticated exploits where you can kind of really follow what was found. And then we have a minimal patch recommendation here to kind of fix this issue. And this is meant to be taken as a starting point, you know, all this context and then put it back into Codex to implement a fix and ship it to production.
Nancy Wang 23:15 ↗
Well, I love that this product design really highlights the simplicity and ease of use.
Alexander Thiemann 23:20 ↗
Yeah, me too. I think it's really very easy and very powerful to really make sure and to kind of scale out the security work that has to be done at every company. So, I'm very excited about this and we're working very closely with customers like DoorDash, Vanta, and Harvey and many more to kind of build this and we hope to roll this out to many more of you and we're hoping that you all sign up on our waitlist.
Nancy Wang 23:45 ↗
And we love feedback. So, thanks so much, Alex.
Alexander Thiemann 23:48 ↗
Yes, thank you.

17 more exchanges in this transcript

Sign in free to read the rest of this interview. No card required.

Sign in to read the full transcript

Cite this transcript

APA, MLA, BibTeX
APA

Brockman, G. (2026, November 12). Cyber at the Frontier: Introducing Daybreak Red Team & Daybreak Defense Factory [Interview transcript]. OpenAI. CEOInterviews.AI. https://ceointerviews.ai/interview/3279333/

MLA

Greg Brockman. "Cyber at the Frontier: Introducing Daybreak Red Team & Daybreak Defense Factory." OpenAI, 12 Nov. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/3279333/.

BibTeX
@misc{brockman2026_3279333,
  author       = {Greg Brockman},
  title        = {Cyber at the Frontier: Introducing Daybreak Red Team \& Daybreak Defense Factory},
  howpublished = {Interview transcript, OpenAI. CEOInterviews.AI},
  year         = {2026},
  month        = {nov},
  url          = {https://ceointerviews.ai/interview/3279333/},
  note         = {Speaker-attributed transcript with timestamps}
}