OpenAI Launches Daybreak Red Team and Defense Factory for Security Testing and Drafting Fixes

OpenAI unveiled two new products on its Daybreak security platform at its Cyber at the Frontier keynote on November 12: Daybreak Red Team and Daybreak Defense Factory, covering attack testing and remediation respectively. President Greg Brockman told the keynote that companies are secure only to the extent of the frontier model they use for defense, and that defense must become continuous.
Daybreak Red Team, now accepting early-access requests, sends coordinated teams of AI agents to penetration-test authorized web applications. The agents analyze application source code and context to explore attack paths and confirm whether a suspected weakness can actually be exploited. They can also validate findings from scanners, bug bounty reports and disclosures, and retest after a fix to confirm it holds. Testing runs in isolated environments, and a review mechanism called Guardian restricts agents to approved applications and endpoints.
Daybreak Defense Factory turns findings into patches, powered by the GPT-6 Cyber model. It connects to GitHub repositories, scans code, traces each vulnerability and prepares a fix; it can also import findings from Snyk, Endor Labs and Wiz, and can run one-off scans, monitor new commits continuously or run on a schedule. Patches are delivered as draft GitHub pull requests, which engineers can inspect, revise with Codex and approve — human review is required before anything is deployed.
GPT-6 Cyber, released on November 5, is the first OpenAI model rated Critical for cybersecurity under the Preparedness Framework. Neither new product has a public price: Red Team is available through a waitlist and Defense Factory as a preview plugin in ChatGPT for workspaces with Daybreak access. For SMEs, the message is that AI defense is moving from point-in-time scanning to a continuous loop — the window between discovery and patch is shrinking, but automated pen testing still needs explicit authorization and scoping, and patches must still pass human review.
Source · Security Boulevard