OpenAI introduced two security products built on AI agents during its "Cyber at the Frontier" livestream on Nov. 12. Daybreak Red Team, an offensive testing tool, launched in early access, while Daybreak Defense Factory hunts for vulnerabilities across a company's code and drafts fixes for them, the company said.
Both sit under Daybreak, the umbrella OpenAI uses for its cyber models and the programs that control who can use them. "We strongly believe that the best defense is a good offense," said Nancy Wang, who leads cyber applied AI at OpenAI, during the stream.
Attack first, then fix
Daybreak Red Team runs OpenAI's own frontier cyber models against a customer's systems. Wang described it as giving "the signal of a superhuman pentest at the scale of automated continuous scanning." A pentest, or penetration test, is an authorized attempt to break into a system to find weak points.
The product tests web applications with and without logins and builds working proof-of-concept exploits, Cyber Press reported. After a fix lands, it reruns those exploits to confirm the hole is closed. OpenAI is taking applications for early access.
Defense Factory handles the other half. It scans a codebase in reproducible cloud development environments, where agents trace a vulnerability to the affected code and use the team's code history to find the right owner, Wang said. Engineers then get "a merge ready pull request with the patch and a test that shows that it works."
When new commits land, the agents review the changes and prepare fixes in the background. Defense Factory also ingests findings from existing scanners as well as from Red Team, Cyber Press reported, and security teams approve every patch before it ships.
What OpenAI says the models found
Clint Gibler, who leads cybersecurity research at OpenAI, said his team used GPT-6 Cyber to find and reproduce 81 previously unknown high-severity vulnerabilities in widely used open-source code. The model also found new attack paths for 30 known issues that had been rated low or medium severity and considered impossible to exploit, he said.
Gibler pointed to the PHP flaw known as PHP2Shell. The bug had existed for 13 years, but its severity wasn't clear until a researcher used OpenAI's latest cyber models to show it could be exploited, he said. OpenAI also used its models to look for bypasses in the patch, work Gibler said would have taken weeks but took a few days.
OpenAI also cited customers. Sophos uses Daybreak agents to investigate threats, and 52% of its managed detection and response cases are now solved end to end by agents, according to the livestream. Average response time fell from 38 minutes to 89 seconds, OpenAI said.
A teased model and its caveats
OpenAI president Greg Brockman showed a chart for an upcoming cyber model that the company is "working on releasing soon" to trusted defenders. Cyber Press reported the unreleased model scored 77.0% on ExploitGym, which tests whether a model can build a working exploit for a known flaw, and 79.1% on PatchBench, which tests fixes that don't break the application.
OpenAI hasn't published those figures in a system card or on its product pages, explainx.ai noted, so they remain livestream claims. Methodology for the chart hasn't been released.
On staying in bounds, OpenAI said GPT-6 Cyber stayed within its authorized scope in 99.9% of evaluated cases without cyber safeguards, and in all cases with production safeguards.
Access stays gated
Access to the underlying models runs through Daybreak's trusted-access program, which asks applicants for identification and details of their professional use, Cyber Press reported. OpenAI calls the capabilities dual-use and concedes the controls can't stop all misuse, according to the same report.
OpenAI said it is also bringing its cyber models to Amazon Bedrock and Google's Gemini Enterprise Agent Platform. It hasn't published pricing for either Daybreak product.
Source: OpenAI, “Cyber at the Frontier”
Additional sources: OpenAI livestream, “Cyber at the Frontier keynote”; Cyber Press, “OpenAI launches AI security agents to find vulnerabilities, test exploits and fix flaws”; explainx.ai, “OpenAI next cyber model: 77% ExploitGym teaser claimed vs verified”
