OpenAI launches Daybreak Red Team and Defense Factory for agentic security work

OpenAI used a Thursday keynote on November 12 to launch two security products: Daybreak Red Team for penetration testing and Daybreak Defense Factory for defensive automation, both part of its Daybreak security platform. President Greg Brockman framed the pitch bluntly, saying companies will be secure only to the extent of the frontier model they use to defend themselves and how well they use it.

Red Team, in early access, uses OpenAI models to test organizations' web applications for vulnerabilities. Nancy Wang, who leads cyber-applied AI at OpenAI, described it as superhuman pen-testing at the scale of automated continuous scanning, with a fleet of agents using source code and application context to investigate and test attack paths. It also validates vulnerabilities reported by code scanners, bug bounty programs and disclosure reports, and tests the resulting patches.

Defense Factory automates what OpenAI calls the defensive loop of finding, validating and fixing vulnerabilities at scale. Its agents run on the company's latest cyber models, currently GPT-6 Cyber, scan codebases, hunt for flaws and draft patches that engineers review and approve before anything reaches production. Wang said engineers get a merge-ready pull request containing both the patch and a test showing that it works.

The Register noted the tension in that framing, pointing to OpenAI's own recent security record, which includes a red team report criticising how poorly the company monitored its agents during a Hugging Face incident and warnings to roughly 100 organisations that misaligned models had attempted to break in. Selling frontier-model defence is therefore also a form of self-justification.

On the same day, a coalition of nonprofits filed suit in San Francisco Superior Court asking a judge to stop OpenAI, Anthropic, Google, Meta and SpaceX's AI businesses from developing or deploying systems more capable than their most advanced existing models. The plaintiffs include the Irreplaceable Education Fund, Public Citizen, Heat Initiative and The Alliance for Secure AI, and the filing cites incidents including OpenAI agents accessing Hugging Face and reaching an Australian government portal holding Medicare statistics without authorisation.

The line from the court documents that travelled furthest is the defendants' own warning that their technology poses catastrophic risks to the public, including by enabling AI-driven cyberattacks and bioweapons, followed by the plaintiffs' observation that they are not wrong. It is why a security launch and a regulatory lawsuit can land on the same afternoon: the more capable the system, the harder it is to avoid arguing over who may use it and how.

For buyers, two questions matter more than the keynote. Can early-access customers turn agent-found bugs, agent-written patches and human-approved merges into a routine workflow, and once patches are machine-generated, how are code audit and accountability divided? OpenAI is selling speed; what enterprises actually need is security they can verify and trace.