OpenAI launched two security tools at its Cyber at the Frontier event on 12 November, sending AI agents to hunt flaws in web apps and write fixes engineers must approve.

The launch lands as attacks get faster and more automated. OpenAI’s argument is blunt: defence has to run on the same class of model as the attack, or it falls behind. Both tools are built for company security teams, not home users, and both stop short of acting alone. A person still reviews every fix before it ships.

Key Points

  • OpenAI launched Daybreak Red Team and Daybreak Defense Factory on 12 November at its Cyber at the Frontier event.
  • Red Team sends teams of AI agents to probe web apps for weaknesses and check whether reported flaws can really be exploited. It is in early access.
  • Defense Factory traces flaws through code and drafts patches, which engineers review and test before merging. It is a preview plugin in ChatGPT.
  • Both run on OpenAI’s GPT-6 Cyber model. OpenAI has published no prices and no independent test results.
  • On the same day, four nonprofits sued OpenAI and four other AI companies in San Francisco, asking a court to halt work on more capable AI systems.

What happened

OpenAI introduced the pair under its Daybreak security platform, which it set up earlier this year to put frontier AI models to work on cyber defence. In August it added Red and Blue access tiers for security researchers. The new tools turn that platform into something companies can buy into, rather than a research programme.

Daybreak Red Team does the attacking. A coordinated team of agents reads an application’s source code and its surrounding context, maps out the ways in, and tries them. Where a scanner or a bug bounty report flags a possible flaw, the agents check whether it can actually be exploited, instead of leaving the team to guess. They also retest patches, to see whether a fix really closes the hole. The work runs in isolated environments, and a Guardian review layer sets the limits on which applications and endpoints the agents may touch.

Daybreak Defense Factory does the repair. It takes findings from Red Team or from tools a team already uses, including Wiz, Snyk and Endor Labs, and connects to the company’s GitHub repositories. Its agents trace each flaw back through the code, work out the cause, and write a patch. Engineers can look at the evidence behind the finding, adjust the patch with the Codex coding agent, and open a draft pull request. Nothing is deployed on its own. The tool also watches new commits and runs scheduled checks, so a flaw introduced next week gets picked up as the code changes.

Both tools run on GPT-6 Cyber, OpenAI’s latest security model, released on 5 November. OpenAI says the model found nearly 1,000 flaws in the Linux kernel, and reports describe it as the first OpenAI model rated Critical for cybersecurity risk under the company’s own Preparedness Framework. The launch follows OpenAI’s release of a free Codex Security scanner on 11 November.

Why it matters

Penetration testing has always been slow and expensive. A human team tests a snapshot of an app, writes a report, and moves on, while the code keeps changing underneath. Continuous agent testing changes the tempo: flaws get found as the code changes, not months later, and every scanner alert gets checked against a real exploit attempt instead of sitting in a queue.

The other side is moving just as fast. The lawsuit filed on the same day cites cases in which AI agents reached systems they should never have touched, including Hugging Face and an Australian government portal holding Medicare statistics. That is the backdrop OpenAI is selling against, and its pitch is that only a frontier model can keep up with a frontier attack.

“We’re now at a point where companies will be secure only to the extent of the frontier model that they use to defend themselves and how well they utilize that,” OpenAI President Greg Brockman said at the keynote.

Nancy Wang, who leads cyber-applied AI at OpenAI, described Red Team as “superhuman pen-test at the scale of automated continuous scanning.” Of Defense Factory, she said: “This way engineers get a merge-ready pull request with the patch and a test that shows that it works.”

“Superhuman” is OpenAI’s word, not a measured result. The company has named no customers, published no prices, and offered no outside test of how often the agents are right or how many false alarms they raise. There is still no independent evaluation of the tools’ detection rate, false positives or patch quality. Companies weighing up the products have only OpenAI’s word so far.

Anthropic is moving in the same direction, putting Claude agents into factory work, while Google’s newest model family is also aimed at coding agents. Not every industry welcomes autonomous agents, though: airlines are already pushing back on AI agents acting on their own.

What happens next

Companies can apply for early access to Daybreak Red Team. OpenAI has not said when it will open the tool more widely, or what either product will cost. Defense Factory is available now as a preview plugin for ChatGPT workspaces that already have Daybreak access.

Three things are worth watching. The first is pricing, which will decide whether this is a tool for large security teams or a wider market. The second is the first customer results, which will show whether agent-written patches hold up in production. The third is the San Francisco lawsuit, which asks the court to freeze development of systems more capable than today’s most advanced models at five companies, including OpenAI.

For now, the tools are a bet that the same models writing the code can also be trusted to break it and fix it, with a human still holding the merge button.

Sources: The Register, SecurityWeek, Neowin, OpenAI. Published 14 November 2026.