OpenAI has turned its Daybreak cyber program into two products. Announced at its Cyber at the Frontier event on November 12, Daybreak Red Team uses teams of AI agents to attack an organization's web applications and confirm which vulnerabilities are actually exploitable, and Daybreak Defense Factory uses agents to trace those bugs through the code and draft patches, SecurityWeek reports.
Why it matters: most security teams aren't short of findings. They're short of time to work out which scanner alerts and bug-bounty reports are real, and then to fix them. These products go after both steps. But they come from the same company selling the model underneath, with no published pricing and no independent data on how often the agents are right.
What each product does
| Daybreak Red Team | Daybreak Defense Factory | |
|---|---|---|
| Job | Pen-test web apps, validate findings | Trace confirmed bugs, draft fixes |
| Inputs | Source code, app context, scanner results, bug-bounty and disclosure reports | Red Team findings, plus imports from Wiz, Snyk and Endor Labs; connected GitHub repos |
| Output | Confirmed, prioritized vulnerabilities; re-tests of patches | Proposed patches; draft GitHub pull requests |
| Safeguards | Isolated environments; "Guardian" review enforces scope | Engineers review and test every patch; nothing deploys automatically |
| Availability | Early access, by application | Preview plugin for workspaces with Daybreak access |
| Price | Not published | Not published |
Red Team's agents use source code and application context to investigate attack paths, according to OpenAI's description quoted by SecurityWeek. Customers decide which applications and endpoints the agents can reach, and OpenAI's Guardian layer enforces those limits. ShiftDelete adds that Red Team can re-run assessments as an app changes and check whether a patch really blocks an exploit.
Defense Factory can scan connected GitHub repositories, watch new commits and run scheduled assessments. Engineers can inspect the evidence behind a finding, adjust the patch with Codex and open a draft pull request. SecurityWeek describes it as a preview plugin in ChatGPT for Daybreak workspaces. Some other reports place it in Codex's plugin marketplace, so check where it appears in your workspace.
The model underneath
ShiftDelete reports that both products run on GPT-6 Cyber, which OpenAI released a week earlier. Fortune reported on November 5 that OpenAI was limiting it to approved Daybreak Red customers first because it could enable attacks if misused; it was trained to refuse less often on tasks that can help attackers as well as defenders. OpenAI says the model found 81 previously unknown high-severity Firefox bugs and produced working exploits for 948 Linux kernel vulnerabilities. Those are OpenAI's numbers, not independent results.
Daybreak itself has two tiers. Daybreak Blue covers defensive work on OpenAI's mainline models. Daybreak Red gives approved organizations its specialized cyber models, per CyberScoop's account of the program's August launch. In September OpenAI said thousands of defenders across 2,000 approved organizations used Daybreak, and it committed $1 billion to subsidized access for utilities, local governments, community banks, nonprofits and open-source maintainers (OpenAI).
What's missing
- Pricing. Neither product has a public price. Red Team's only option is a request for early access.
- Accuracy data. OpenAI hasn't published false-positive rates or how often the drafted patches are correct, ShiftDelete notes. That matters: CyberScoop has reported research finding that strong models can struggle to fully patch a bug, or can add new bugs with their fixes.
- Data handling. Usage limits, data retention terms and what happens to your source code aren't described in the coverage.
Who should care, and what to ask
Application security teams already buried in scanner and bug-bounty triage are the obvious audience, especially those on Wiz, Snyk or Endor Labs, whose findings Defense Factory can import. Before you apply, ask OpenAI four things. How is Guardian's scope enforced, and is it logged? Where do your code and test data go, and how long are they kept? What validation rates has OpenAI measured on real customer apps? And how is a "confirmed" finding shown to you, with a reproducible proof or just the agent's verdict?
Treat the drafted patches like pull requests from a new contractor: review them and test them, and don't assume a passing re-test means the whole bug class is gone.
Sources
- SecurityWeek — OpenAI Unveils Daybreak Red Team, Defense Factory Security Tools securityweek.com
- ShiftDelete.Net — OpenAI Daybreak Red Team and Defense Factory Explained shiftdelete.net
- Fortune — OpenAI debuts new cybersecurity-focused AI model amid growing alarm about AI-driven cyber attacks fortune.com
- OpenAI — Daybreak for frontline defenders openai.com
- CyberScoop — OpenAI launches Daybreak Blue and Daybreak Red cyber services (Aug. 10) cyberscoop.com