On 12 November 2026, at its Cyber at the Frontier event, OpenAI launched Daybreak Red Team and Daybreak Defense Factory. Red Team is a set of AI agents that pentest web applications, check whether a finding is actually exploitable, and retest a fix.
We've been reading the write-ups since it landed. A lab the size of OpenAI shipping this is, frankly, good for the category. It tells founders that continuous testing by agents is a real product, not a pitch deck. We've been building that product as our only product. For OpenAI, Daybreak sits inside a much larger lab. That difference matters more than the launch-day headlines.
If you need to start this month, or the paths you worry about don't stay inside one web app, HOG3 is the one we'd use.
What OpenAI announced
Daybreak is the umbrella program OpenAI already had for cybersecurity. The 12 November event added two products to it. (OpenAI Daybreak; Cyber at the Frontier)
OpenAI describes Daybreak Red Team as "AI penetration testing for web applications." A team of agents tests those apps, confirms vulnerabilities, and checks whether a patch stops the exploit. SecurityWeek, The Register and Neowin all say it's in early access. You can register interest, though the form is clear that registration doesn't guarantee access. (OpenAI Daybreak Red Team; Register interest; SecurityWeek, 13 November 2026; The Register, 12 November 2026)
Defense Factory sits next to it. Agents look at findings and draft patches. SecurityWeek describes it as a preview plugin in ChatGPT, and only for workspaces that already have Daybreak access. OpenAI's own setup guide also calls it a preview. (OpenAI Daybreak Defense Factory; Set up Daybreak Defense Factory)
superhuman pen-test at the scale of automated continuous scanning
Nancy Wang, who leads cyber-applied AI at OpenAI, as reported by The Register, 12 November 2026
That's the pitch, in her words. Founders still have to read the pages and see what is actually on offer.
What Daybreak Red Team actually does
We've seen a lot of product pages that hide the scope. OpenAI's isn't one of them. It's short, and it tells you what the agents do.
They use "source code and app context to investigate attack paths and confirm vulnerabilities." That's useful context for a web application. It is also a narrower job than following a path from an app, through an API, into a cloud role or an identity.
They can also take findings you already have, from code scanners, bug bounty submissions, and disclosure reports, and test those. After you patch, they can "test whether a patch stops the exploit" and "repeat the assessment as your apps change." That's a way to work through a web-app backlog, and to retest a fix on that same surface.
Scope is supposed to stay inside a boundary you set, with isolated environments and something OpenAI calls Guardian review. You control which apps and endpoints the agents can reach. They don't publish a proper technical write-up of Guardian on the product page, so we won't invent one. (OpenAI Daybreak Red Team; SecurityWeek, 13 November 2026)
The Register says Red Team uses OpenAI models. What nobody has published, on the official page or in the launch coverage we read, is a price, a date for general availability, or any customer results.
What the public sources do not claim
Red Team isn't generally available. The reporting calls it early access, and OpenAI's form says interest isn't a promise of access.
The official scope is web applications. The product page doesn't claim to pentest APIs, cloud infrastructure, or identity systems.
And the name will trip people up. Daybreak Red Team is not the same thing as Daybreak Red. Daybreak Red is an access tier in OpenAI's Trusted Access for Cyber program. Daybreak Blue is the defensive counterpart. Daybreak Red Team is the November product. Older forum threads about "getting Daybreak Red" are usually about model access, not this. (OpenAI help: Trusted Access for Cyber)
We also couldn't find, in any launch piece we could fully read, benchmark numbers, design-partner names, or a description of where customer source code sits during a Red Team run.
None of that makes it a bad product. A focused web-app test, with confirmation and a retest, is real work. It is also one surface. You need to know which surfaces are in the test you're buying, and which are not.
Defense Factory, and the engineer who still has to say yes
Defense Factory is a companion product, not the pentest itself. It sits in the developer workflow around a web application: findings in, a draft change out, and an engineer still in the middle.
OpenAI says agents "trace vulnerabilities through your code and develop patches," and that "engineers review and test proposed patches before merge." SecurityWeek is blunt about the limit: patches are not deployed automatically. In The Register's account, Wang said engineers get "a merge-ready pull request with the patch and a test that shows that it works." (OpenAI Daybreak Defense Factory)
Findings can come from Red Team or from tools you already run. SecurityWeek reports imports from Wiz, Snyk and Endor Labs, and OpenAI's setup guide lists the same three names. It can scan connected GitHub repositories, watch new commits, and run on a schedule. Engineers can look at the evidence, tweak a patch with Codex, and open a draft pull request.
A couple of setup details in OpenAI's own docs are easy to skim past. Cloud scans run in Codex cloud, and Defense Factory doesn't support workspace-level GitHub connections. Each person who starts a scan needs their own personal GitHub connection.
OpenAI's developers blog, which still calls this a research preview, also says a scan may take several hours on a larger repository. (Set up Daybreak Defense Factory; Scaling cyber defenders with Daybreak)
We don't claim that workflow. We write remediation guidance for your engineers. Your team makes the change. We don't touch your code or production, and we don't open a merge-ready GitHub pull request. That's a different job from testing how an attacker would move through your systems. (About HOG3)
How HOG3 compares
HOG3 is a continuous autonomous AI pentesting platform, and it's a public offer today. You don't have to join a waitlist to talk to us. This is our product, not a side program inside a larger lab.
What we want to compare is practical: who can start, what actually gets tested, and what a finding is worth when it lands on an engineer's desk.
Why HOG3 is faster to put to work
We're not going to invent a number for how quickly a run finishes. When we say faster, we mean time to start, and whether you can keep testing as you ship. You can check both of those on the public pages.
You don't have to sit on a waitlist. HOG3 is live. We start with a demo, agree the systems and the boundaries, and nothing is tested until you authorize it. If it fits, a scoped free pentest may be available by agreement before a paid subscription.
Daybreak Red Team isn't generally available. The coverage we read calls it early access, the interest form doesn't guarantee you'll get in, and nobody has published a date for general availability. (HOG3 pricing; Register interest)
You can also keep testing as you ship. Our assessments can run daily, weekly, monthly, or on a custom cadence, and again on releases, infrastructure changes, and fixes. Daybreak Red Team can "repeat the assessment as your apps change," which is the right idea for a web app. The product itself is still gated. You can't run a cadence you don't have access to yet. (HOG3)
Why the coverage is stronger
We've watched enough incidents to be wary of a test that stops at the web app. A feature can work exactly as it was built. The next one can too. Put them together and you sometimes get a path into an account, or into data, that nobody meant to open. Those chains often run through an old token, a cloud role, or an identity setting that hasn't been looked at in a year.
That's the problem we built HOG3 around. The public offer covers applications, APIs, cloud infrastructure, mobile applications, identity systems, and external attack surfaces, including what sits behind a login. Authenticated application and API testing is part of it.
We don't leave findings as a pile of isolated mediums. We link them into a working path. Each high or critical finding comes with evidence you can reproduce, and with some sense of what it means for the business. A person validates what we find. We only close a finding when the same path is tried again and no longer works. (HOG3; Three mediums are a critical; What a proof contains)
Daybreak Red Team, on its official page, tests web applications. Source-code context, scanner validation, and draft pull requests are part of that web-app workflow. They don't turn a single-surface test into coverage of APIs, cloud, identity, mobile, or the external surface.
Why a specialist beats a side product
OpenAI builds models, consumer products, and a lot else besides. Daybreak is one program inside that. HOG3 is a continuous autonomous AI pentesting platform. That's the company. When a finding, a scope question, or a retest lands, it's the work our product exists to do.
We publish a roadmap for that product. What's live today is continuous autonomous pentesting: mapping the surface, chaining findings, and proving impact. The next work on that page is about helping your team close a finding faster, then retesting it. We won't sell you the planned items as if they were already in the offer.
Because this is the job, the controls around it are part of the product, not a bolt-on. Nothing is tested until you agree the scope and authorize the work. A person validates what the platform finds. We only close a finding when the same path is tried again and no longer works.
Assessments can run daily, weekly, monthly, or on a custom cadence, and again on releases, infrastructure changes, and fixes. That's the difference between a tool you might get access to and a testing program you can actually run. (HOG3; HOG3 pricing; HOG3 roadmap)
Side by side
This table is only what both sides have published. If we don't have a public claim for a cell, we say so.
| Factor | HOG3 | Daybreak Red Team |
|---|---|---|
| Availability | Available now. Start with a demo. A scoped free pentest may be available by agreement. | Early access, according to SecurityWeek, The Register and Neowin. Registration does not guarantee access. No public price or general-availability date. |
| Surfaces | Applications, APIs, cloud infrastructure, mobile applications, identity systems, and external attack surfaces, including authenticated apps and APIs, as agreed in scope. | Official page: web applications. It does not claim APIs, cloud, or identity. |
| Attack paths | Findings linked into a working path across those surfaces, the way an attacker would move. | Agents use source code and app context to investigate attack paths and confirm vulnerabilities in the web apps in scope. |
| Cadence | Daily, weekly, monthly, or custom, plus tests on releases, infrastructure changes, and fixes. You can start that cadence now. | Can repeat the assessment as apps change, once you have access. The product is still in early access. |
| Proof | Validated findings with reproducible evidence and business-impact context. A person validates what HOG3 finds. | Confirms whether a vulnerability is exploitable in the web apps in scope, including findings from scanners, bug bounties, and disclosure reports. |
| Retest | A finding closes only when the same path is tried again and no longer works. | Tests whether a patch stops the exploit. |
| Controls | Nothing is tested until you agree the scope and authorize the work. A person validates what HOG3 finds. | Isolated environments and Guardian review. You control which apps and endpoints agents can reach. |
| Fixes | Remediation guidance for your engineers. HOG3 does not change your code or production systems. | Defense Factory, a companion product, can draft a merge-ready pull request for a web-app finding. An engineer reviews and tests it before merge. Patches are not deployed automatically. |
| Alongside other tools | Can work alongside your security team, scanners, and an existing pentesting provider. | Defense Factory can import findings from Wiz, Snyk, and Endor Labs, and from Red Team. |
We're not claiming a formal integration with Daybreak. We can sit next to tools you already use. That's as far as our public position goes.
Who should use what
If you need to start now, you want testing that keeps up as you ship, or the paths you actually worry about run through APIs, cloud, identity, or more than one surface at once, that's the work we built HOG3 to do. It's also the work we do every day, because it's the product.
If what you want is a draft pull request against one web app, and you already have Daybreak access, Defense Factory is built around that workflow. That's a different job from following an attacker across your estate.
You could also run both. We already work alongside existing providers. Just check the scopes so you aren't paying twice for the same surface.
And if you only run a scheduled pentest a few times a year, neither of these is a reason to throw that assessment away. It also isn't a reason to treat the weeks in between as untested. We wrote about that gap in What happens between scheduled pentests?.
A short checklist
You don't have to buy anything this week. A few questions will still help, whether you talk to OpenAI, to us, or to both.
- Write down the surfaces you actually care about: the web apps, the APIs, the cloud accounts, and the identity systems. Then ask any vendor which of those it tests. For Daybreak Red Team, start from what OpenAI has published, which is web applications.
- Ask how a finding is confirmed. A scanner item, a bounty report, and a demonstrated path are not the same thing.
- Ask what happens after a patch. You want the same path, tried again, and a recorded failure.
- Confirm the rules of engagement. Who authorizes the test, which environments can be reached, and who reviews the output.
- If you look at Defense Factory, keep the human review. Patches are not deployed automatically. You'll need a Daybreak-enabled workspace, a personal GitHub connection for each person who starts a scan, and some patience if the repo is large. A scan can take hours.
- Don't treat a registration form as access, and don't mix up Daybreak Red with Daybreak Red Team.
Questions founders actually ask
What is OpenAI Daybreak Red Team? It's OpenAI's AI penetration-testing product for web applications, announced on 12 November 2026 at the Cyber at the Frontier event. A team of agents uses source code and application context to investigate attack paths, confirm whether a vulnerability is exploitable, validate findings from scanners, bug bounties and disclosure reports, and retest whether a patch stops the exploit.
Can we use it today? Not as a generally available product, based on what we've read. SecurityWeek, The Register and Neowin describe Red Team as early access. You can register interest with OpenAI, and the form says that doesn't guarantee access. Defense Factory is a preview plugin in ChatGPT for workspaces that already have Daybreak access. We didn't find a public price or a general-availability date.
Does Defense Factory automatically merge or deploy patches? No. OpenAI says engineers review and test proposed patches before merge. SecurityWeek says patches are not deployed automatically. You get a merge-ready pull request. A person still has to accept it.
Is Daybreak Red Team the same as Daybreak Red? No. Daybreak Red is an access tier for approved, authorized red-team and research work on OpenAI's models. Daybreak Red Team is the November penetration-testing product. It's worth using the full name, because people are already mixing them up.
Does Daybreak Red Team test APIs, cloud and identity, or only web apps? The official page describes penetration testing for web applications. It doesn't claim coverage of APIs, cloud infrastructure, or identity systems. If those surfaces matter in your environment, treat them as out of scope for this product until a primary source says otherwise.
How does HOG3 compare to OpenAI Daybreak Red Team? HOG3 is ahead on the decisions a founder has to make this month. We're available now. We run continuous autonomous AI pentesting across applications, APIs, cloud infrastructure, mobile applications, identity systems, and external attack surfaces, we chain findings across those surfaces the way an attacker would, and we only close a finding after a retest. A person validates what we find, and nothing runs until you authorize the scope. Daybreak Red Team is an early-access OpenAI product for pentesting web applications. It can use source-code context, check a scanner or bug-bounty item, and, through Defense Factory, draft a pull request an engineer still has to accept. There's no public price or general-availability date, and the official scope is web apps. If you need to start now, or your risk isn't confined to a single web app, HOG3 is the product we'd use.
How is this different from a vulnerability scanner? A scanner gives you a list of possible weaknesses. OpenAI's stated aim for Red Team, and what we offer at HOG3, is to confirm whether a path is actually exploitable, then retest the fix. You're trying to buy confirmation and closure, not a longer backlog.
If you want to see HOG3 against your own systems, start at hog3.com. The security documentation is on the Trust Center.
Sources
- OpenAI, Daybreak Red Team
- OpenAI, Daybreak Defense Factory
- OpenAI, Daybreak
- OpenAI, Cyber at the Frontier
- OpenAI, Register interest in Daybreak Red Team
- OpenAI, Set up Daybreak Defense Factory
- OpenAI Developers, Scaling cyber defenders with Daybreak
- OpenAI help, Daybreak Trusted Access for Cyber overview
- SecurityWeek, 13 November 2026
- The Register, 12 November 2026
- Neowin, 13 November 2026
- HOG3
- HOG3 About
- HOG3 pricing
- HOG3 roadmap
- HOG3 Trust Center