OpenAI Turned Its Cyber Models Into a Pentest Product. Red Team Is Early Access With No Price, and the 77% Model on the Slide Has No Name.
At its Cyber at the Frontier livestream on November 12, OpenAI launched Daybreak Red Team, an early-access team of agents that pentests web applications from source code and app context, validates scanner and bug-bounty findings, and rechecks whether patches hold inside isolated environments with Guardian review limiting scope, and Daybreak Defense Factory, a preview ChatGPT plugin for Daybreak workspaces that scans connected GitHub repos, traces findings, and drafts patches and tests as draft pull requests that engineers merge. Greg Brockman referred to a billion-dollar Daybreak defense fund. The livestream also showed an unnamed cyber model at 77.0% on ExploitGym and 79.1% on PatchBench, numbers that exist only on a slide; the only documented score is GPT-6 Cyber's 48.5%. No price, no GA date. We read OpenAI's pages and the transcript.
By T.W. Ghost
OpenAI used its "Cyber at the Frontier" livestream today, November 12, 2026, to launch two products under the Daybreak name, rather than the model many viewers expected. Daybreak Red Team, "launching in early access," is a team of agents that pentests web applications from source code and application context, validates findings that come in from scanners and bug-bounty programs, and rechecks whether a patch held. It runs inside isolated environments, with what OpenAI calls Guardian review limiting scope. Daybreak Defense Factory, a preview ChatGPT plugin for workspaces with Daybreak access, imports findings from Red Team, Wiz, Snyk, and Endor Labs, scans connected GitHub repositories and monitors commits, traces each finding, and drafts the patch and its tests as a draft pull request. An engineer reviews, edits with Codex if needed, and merges. Nothing deploys on its own. Greg Brockman referred to a billion-dollar Daybreak defense fund for frontline defenders.
The number everyone repeated came from a slide. OpenAI showed an unreleased cyber model at 77.0% on ExploitGym and 79.1% on PatchBench, and the press set it against GPT-6 Cyber's published 48.5%. The slide model has no name, no system card, no access path, and no stated track, standard or hardened, so the comparison is not like for like and we do not make it. Neither product has a price or a GA date. We read OpenAI's event and Red Team pages, the livestream transcript, and SecurityWeek. Here is what held up.
What OpenAI Launched
| Daybreak Red Team | Daybreak Defense Factory | |
|---|---|---|
| Status | Early access | Preview ChatGPT plugin for workspaces with Daybreak access (press) |
| Does | Pentests web apps from source and app context; validates scanner and bug-bounty findings; rechecks fixes | Imports findings from Red Team, Wiz, Snyk, Endor Labs; scans GitHub repos and monitors commits; traces findings; drafts patches and tests as draft PRs |
| Controls | Isolated environments; Guardian review limits scope (vendor) | Engineers review, edit with Codex, create draft PRs; no automatic deploy |
| Price; GA | Not published | Not published |
The Benchmarks, Sorted by Who Reported Them
| Claim | Number | Who |
|---|---|---|
| GPT-6 Cyber on ExploitGym; GPT-5.6 Cyber | 48.5%; 32.1% | Vendor (Nov 5 launch post) |
| Unreleased cyber model on ExploitGym; PatchBench | 77.0%; 79.1% | Unverified (livestream slide via press and X; no OpenAI post, no model name, no card) |
| Anthropic Mythos on the same slide | About 38% | Unverified (Gizmodo Japan via explainx.ai) |
| Daybreak defense fund | "Billion dollar" | Vendor statement (livestream) |
Three honest observations:
- The 77% is a slide. Until OpenAI posts a system card or a model name, it is a claim about a model nobody can call, on a track nobody can name.
- Scope control is OpenAI's own gate. Guardian review is described by the vendor; no third-party audit is cited. A pentest agent with source-code access is exactly the thing you want someone else to have checked.
- The human merge is the honest part. Defense Factory drafts; it does not ship. That limits the "automated defense" framing, and it is the right limit.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Red Team and Defense Factory announced Nov 12, 2026 at Cyber at the Frontier | Verified (OpenAI pages; transcript) |
| Red Team early access; capabilities; isolated environments; Guardian review | Verified (vendor) |
| Defense Factory capabilities; draft PRs; no automatic deploy | Verified (vendor, via SecurityWeek quotes) |
| Defense Factory as a preview plugin for Daybreak workspaces | Press (SecurityWeek) |
| Pricing; GA date | Not published |
| Defense fund | Vendor statement |
| 77.0% ExploitGym; 79.1% PatchBench | Unverified; held |
| Mythos about 38% | Unverified; held |
Who Should Use It
Ask for Red Team access if you run web applications with source you can hand over and a scanner or bug-bounty queue that needs validating, and you can wait on a list with no price attached.
Skip the 77% in any deck, and skip Defense Factory unless you already have Daybreak access; the patch it drafts still needs your engineer's merge, which is the point.
The frontier is elsewhere: the only cyber scores OpenAI has put in writing are GPT-6 Cyber's, and the only large-scale result is the 948 root exploits on Linux. Our model comparison has the current standings side by side.
Sources
- OpenAI: Cyber at the Frontier
- OpenAI: Daybreak Red Team
- YouTube: Cyber at the Frontier livestream
- SecurityWeek: OpenAI unveils Daybreak Red Team, Defense Factory security tools
- explainx.ai: OpenAI's next cyber model at 77% on ExploitGym, claimed vs verified
- X: trending summary of the livestream
*Not sure whether a pentest agent fits your security stack? Take the free 2-minute quiz and get matched. Then read about the model whose score is the only one OpenAI has documented: GPT-6 Cyber.*