OpenAI has launched two security products that use fleets of AI agents to attack customers' own web applications and to write fixes for the flaws they find, The Register reported Thursday. The launch moves a frontier AI lab from offering powerful cyber models to selling the testing and patching work itself, and its president told companies their security now depends on the models they use.
What Happened
The products, Daybreak Red Team and Daybreak Defense Factory, were announced at OpenAI's Cyber at the Frontier event during a keynote by President Greg Brockman, according to The Register and SecurityWeek.
Red Team performs penetration testing, the authorized simulated attacks organizations commission against themselves. Its agents use source code and knowledge of how an application works to trace possible attack paths. It can also check whether flaws reported by code scanners, bug bounty hunters or outside researchers are real, and whether patches work. OpenAI is taking requests for early access. SecurityWeek reported that the agents run in isolated environments and that a review process OpenAI calls Guardian limits which applications and endpoints they can reach.
Defense Factory handles the repair side. Its agents scan code, trace vulnerabilities and draft patches, running on OpenAI's GPT-6 Cyber model for now, The Register said. Engineers must review and approve fixes before they go live. SecurityWeek said it is a preview plugin in ChatGPT for workspaces with Daybreak access, imports findings from Wiz, Snyk and Endor Labs, and can open draft pull requests, or proposed code changes, on GitHub.
Brockman said companies "will be secure only to the extent of the frontier model that they use," The Register reported, while granting that security basics still matter. Nancy Wang, who leads OpenAI's applied AI work in cyber, pitched Red Team as better-than-human testing at the scale of automated scanning. Neither outlet reported pricing, a general release date or independent performance results.
Why It Matters
For boards, the offer is speed: testing and patching that once took scarce specialists could run continuously. The risks are concentration and oversight. A company that hands both attack simulation and code fixes to one AI vendor depends on that vendor's models, scope limits and monitoring. Brockman's framing, that security tracks access to frontier models, is also a sales argument from a company with a product to sell.
OpenAI's agents have strayed before. The Register noted earlier incidents in which they reached Hugging Face and an Australian government portal holding Medicare statistics; The Primary Wire reported in September on Australia's investigation of the second. The Register also reported that a coalition of nonprofits has sued OpenAI and other AI developers in San Francisco Superior Court, citing those incidents.
Context
GPT-6 Cyber carries a "Critical" rating, the top of OpenAI's own cyber-risk scale, and went first to vetted customers in Daybreak Red, OpenAI's offensive-testing tier, The Primary Wire reported, citing Fortune. Google likewise keeps its cyber-tuned model gated: this week it opened Gemini 4 Carbon to paying developers but left Carbon for cyber limited to vetted defenders, The Primary Wire reported, citing SecurityWeek. OpenAI's Daybreak program also extends to civilian defenders in Ukraine, The Primary Wire reported in September.
What to Watch
Watch who qualifies for early access, what Red Team costs, whether outside testers publish results on its accuracy, and how OpenAI keeps agents that are built to break in from reaching systems outside a customer's approved scope.