OpenAI introduced Daybreak Red Team and Daybreak Defense Factory on 12 November as part of its Daybreak security platform, LavX reported. Red Team, released in early access, tests web applications for vulnerabilities. Defense Factory scans code and prepares repairs for engineers to approve before they reach production. Together, the products move OpenAI’s cyber models into the work of finding and fixing weaknesses in software.
Key points
- Daybreak Red Team uses OpenAI models to test web applications and check reported vulnerabilities.
- Daybreak Defense Factory uses agents to identify flaws and draft patches, with engineers approving fixes before production.
- Defense Factory includes GPT-6 Cyber, which had previously reached select approved Daybreak Red customers.
Daybreak Red Team checks reported flaws
Nancy Wang, who leads cyber-applied AI at OpenAI, described Red Team as a continuous scanning service, LavX reported. Its agents examine a web application’s source code, the context in which the application runs and possible routes an attacker could take. That gives the service more to work with than an isolated warning in a scan: it can examine a suspected weakness in the application where the weakness might be used.
Checking a website for security flaws would, under that approach, involve examining its code and possible routes of attack repeatedly. A reported flaw could be checked against the application, and a proposed repair could be tested before it is approved.
Red Team is also intended to help security teams check findings from code scanners, bug bounty programmes and vulnerability disclosure reports. Those are different ways a possible flaw can reach a team, from an automated warning to a report from someone who has examined the application. OpenAI’s product tests patches as well, putting a proposed repair back through a security check rather than treating the preparation of a fix as the end of the work.
The early-access release puts that testing service alongside Defense Factory, whose work begins with code that may need repairing. The two products cover different parts of the same process: investigating whether a weakness can be found or exploited, and preparing code changes to address one. The descriptions concern what the tools are designed to do, rather than results from a measured deployment of either product.
Daybreak Defense Factory prepares patches for review
Defense Factory uses agents built on OpenAI cyber models, including GPT-6 Cyber, to scan codebases, identify vulnerabilities and draft patches. A codebase is the collection of files from which an application is built. The proposed change is packaged as a pull request — a way of presenting a code change for review before it is incorporated into the software.
Wang said engineers receive a pull request containing both the patch and a test intended to confirm that the fix works, LavX reported. That pairing matters because a change to code and a check of that change are separate tasks. The test gives the reviewer something to run and inspect alongside the proposed repair, while engineers retain the decision on whether each fix should go into production.
For companies building on the tools, the point of handover is therefore a proposed code change, not an automatically deployed repair. Security teams could use Red Team to investigate incoming reports and test patches, while engineering teams would assess Defense Factory’s proposed fixes and their accompanying tests. In the workflow OpenAI describes, the agents do the scanning and drafting, and engineers approve each change before production.
GPT-6 Cyber reaches the Daybreak workflow
The model named in Defense Factory had already been introduced to select approved Daybreak Red customers through the API and Codex on 5 November, TestingCatalog reported. AI Affairs reported its launch with a Critical cybersecurity rating. The new products put model capabilities into services organised around application testing and the review of proposed repairs.
On OpenAI’s ExploitGym evaluation, GPT-6 Cyber scored 48.5%, against 32.1% for GPT-5.6 Cyber, according to the results reported by TestingCatalog. In black-box web penetration testing, where the model tests a running application rather than working from its source code, the reported vulnerability recall was 40.3%, against 32.0% for GPT-6 Astra. Those figures concern particular tests of models; Red Team and Defense Factory are products that also organise how findings and fixes reach security teams and engineers.
OpenAI also evaluated whether the model stayed within boundaries set for authorised testing. In ExploitGym Honeypot, GPT-6 Cyber avoided out-of-scope services in 99.93% of tests without a deployment safety prompt and stayed within bounds in 100% with that prompt, according to TestingCatalog. OpenAI also accompanies the model’s introduction with cyber-focused activity tracking, screening for malicious use, safeguards for login secrets, access restricted to approved IP addresses, isolated execution environments and human supervision.
