OpenAI Launches Two Security Products at Cyber at the Frontier: Daybreak Red Team Uses Agent Teams to Pen-Test Web Apps and Retest Patches (Early Access), and Daybreak Defense Factory Scans Connected GitHub Repos and Prepares Fixes as a Plugin (Research Preview)

Background and the problem

On November 4 and 5, OpenAI published research in which a new model validated 948 local privilege escalation bugs in the Linux kernel, then introduced GPT-6 Cyber for approved Daybreak Red customers, and announced a Cyber at the Frontier livestream for November 12 at 10am PT. Security teams have long struggled with scanners that report far more issues than are actually exploitable, while fixes wait in a queue. This keynote packages frontier models' bug-finding ability into two enterprise products, one for attack and one for repair.

Architecture and how it works

Daybreak Red Team is a coordinated team of agents that reads source code and app context to work through attack paths and confirm vulnerabilities. It can also check findings from code scanners, bug bounties and disclosure reports one by one, and retest whether a patch actually stops the exploit. Operators set which apps and endpoints the agents can reach, testing runs in isolated environments, and Guardian review applies. Nancy Wang, who leads cyber applied AI, called it "superhuman pen-test at the scale of automated continuous scanning." Daybreak Defense Factory ships as a ChatGPT plugin and runs scans in Codex cloud: it traces a vulnerability through the code, tries to reproduce it in an isolated environment where practical, and prepares a patch with a test. Patches are never merged automatically; engineers review them, refine them with Codex, or open a GitHub draft PR. Red Team findings can feed straight into Defense Factory.

Benchmarks and measured results

OpenAI has published no accuracy or false-positive figures for either product. The verifiable model numbers still come from the November 5 GPT-6 Cyber system card: a 48.45% intended-exploit success rate on ExploitGym versus 42.43% for GPT-6 Astra (vendor-reported). A keynote slide showed an unnamed "Upcoming Cyber Model", and several viewers posted screenshots citing a higher ExploitGym score, but OpenAI has released no written material on it, so we do not report that number.

Getting started

For Defense Factory, open Plugins in a ChatGPT workspace with Daybreak access, search for Daybreak Defense Factory (Preview), install and enable it, open Daybreak in the sidebar, choose Run an agentic scan, and connect GitHub with a personal account (workspace connections are not supported). Start with one folder such as src/; a large repository can take several hours. Cloud scans are billed by token usage outside the plan's included allowance, so check the Usage page first. Red Team requires an early access request at openai.com/daybreak.