OpenAI Daybreak: AI agents hack your web app, then write the patches

What's new

OpenAI is pushing further into IT security. At its "Cyber at the Frontier" event on November 12, 2026, the company unveiled two new products under the Daybreak banner. Daybreak is OpenAI's initiative to put advanced AI models to work on authorized security tasks.

• Daybreak Red Team runs test attacks on web applications to find and confirm vulnerabilities.
• Daybreak Defense Factory analyzes vulnerabilities that have been found and prepares fixes in the code.

Both tools rely on AI agents, meaning AI systems that work through multiple steps on their own. They're aimed at companies and their security and engineering teams, not consumers. You can currently apply for early access to Red Team. Defense Factory is available as a preview plugin in ChatGPT workspaces that already have Daybreak access. Reports from SecurityWeek and other trade outlets don't mention pricing or a date for general availability.

Daybreak Red Team: automated break-in testing

A penetration test, or pentest, is a controlled attack on your own systems. Experts try to break in, with permission, before real attackers do. These tests matter, but they take a lot of time and expertise. That's where Red Team comes in.

According to OpenAI, a group of coordinated agents works together. They read an application's source code and context, look for possible attack paths and check whether a vulnerability can actually be exploited. Red Team can also validate findings that come from other sources:

• alerts from code scanners, programs that automatically comb code for common bugs,
• reports from bug bounty programs, which pay outside hackers for discovered flaws,
• reports from vulnerability disclosure programs.

That's useful because many of these reports are false alarms or overrated. Red Team is meant to provide evidence of which flaws can really be reproduced, so teams can better decide what to fix first. According to Help Net Security and Security Boulevard, Red Team can also attack a finished patch again. That way you can verify that the fix actually stops the attack instead of just assuming it does.

To keep the agents from going after systems that are off-limits, they run in isolated environments. According to OpenAI, a control mechanism called Guardian restricts which applications and endpoints the agents can reach. Companies define these boundaries themselves.

Daybreak Defense Factory: from alert to patch

Defense Factory handles the repair work. A security alert alone doesn't tell developers how a flaw affects their application or how to fix it. So the agents trace a vulnerability through the code, look for the root cause and prepare a proposed fix.

The starting point is findings from Red Team or from existing security tools. According to SecurityWeek, Defense Factory pulls in vulnerability data from Wiz, Snyk and Endor Labs, among others. Help Net Security describes the workflow in more detail:

• You install the "Daybreak Defense Factory (Preview)" plugin in ChatGPT and connect a GitHub repository.
• Scans run in the Codex cloud. Everyone who starts a scan needs their own GitHub connection.
• You can scan the default branch of an entire repository or just selected folders and files.
• Teams can monitor new commits, meaning newly pushed code changes, on the default branch, schedule regular scans and have pull requests checked for security before merging.
• In the desktop app, you can also check local copies and uncommitted changes. An optional deep scan mode goes through the code multiple times, which takes longer and uses more tokens, i.e. more compute.

Engineers can use Codex, OpenAI's coding agent, to view the evidence for a finding, add context and refine the proposal. The end result is a draft GitHub pull request, a proposed change the team can accept or reject.

Humans stay in the loop

OpenAI stresses repeatedly that patches are not rolled out automatically. Humans have to review and test them before they're merged into the code. According to the official announcement, the tools are meant to support security professionals, not replace them. Korean tech site Hitech notes that this approval step is likely to be a key checkpoint for companies. The agents get access to source code and repositories, so companies should set clear rules for access rights and logging beforehand.

Brockman's message and the bigger picture

In his keynote, OpenAI President Greg Brockman said, in essence, that companies will only be as secure as the most advanced AI models they use for defense, and how well they use them. Classic security methods remain important, he said, but given the models' growing capabilities, companies need to rethink their processes. The Register put it bluntly in its headline: OpenAI is telling organizations they're doomed if they don't use its products.

The backdrop is an arms race: AI is finding security flaws faster and faster. Around the same time, for example, SecurityWeek reported that a critical flaw in the Apache HTTP Server that let attackers execute code remotely was found and patched using tools from Anthropic and Google. The faster flaws surface, the more fixing them becomes the bottleneck. That's exactly what Defense Factory targets. The flip side: the same capabilities are available to attackers, who can automate vulnerability hunting and attack steps.

Our take

With Daybreak, OpenAI is going head-to-head with established security vendors, and with Anthropic and Google, which also use their models to hunt for vulnerabilities. What's new is mainly that finding, confirming, fixing and re-testing come together in a single workflow. According to Techgenyz, this could be especially appealing for smaller teams without in-house pentesters looking to work through a backlog of security reports. But it doesn't replace expertise.

A lot is still unclear. There are no independent tests of how reliably the agents find flaws or how good their patches are. Pricing and general availability are unknown. Agents that carry out attacks and access code repositories are also a risk in their own right. How well the Guardian boundaries hold up in practice will only become clear in real-world use. For now, Daybreak Red Team and Defense Factory are tools for enterprise customers with access, so you can't easily try them out yourself.