OpenAI is moving into two jobs that security teams have long done by hand: proving a vulnerability is real, and writing the fix. At its Cyber at the Frontier event on Nov. 12, the company announced two products under its Daybreak name. Daybreak Red Team uses coordinated AI agents to investigate attack paths in web applications. Daybreak Defense Factory uses agents to trace vulnerabilities through a codebase and prepare patches. SecurityWeek and Help Net Security both reported the launch on Nov. 13.

The two tools are built to work together. Red Team finds and confirms weaknesses. Defense Factory turns confirmed findings into proposed code changes. Then Red Team can be pointed back at a patch to check whether it actually blocks the exploit. Under OpenAI's design, engineers review and test proposed patches before they are merged, and SecurityWeek notes that patches are not deployed automatically.

Red Team: agents that check whether a finding is real

Every security team knows the triage problem. Code scanners produce long lists of possible issues, many of them false positives. Bug bounty programs and vulnerability disclosure inboxes bring in reports of very uneven quality. Someone has to read each one, reproduce it and decide whether it matters.

Daybreak Red Team is aimed at that bottleneck. According to SecurityWeek, the product has multiple agents identify and investigate weaknesses in web applications. "Agents use source code and app context to investigate attack paths and confirm vulnerabilities," SecurityWeek quoted OpenAI as saying. The tool also examines findings from outside sources, including code scanner output, bug bounty submissions and vulnerability disclosure reports, so it acts as a validation layer as well as a hunter.

An "attack path" is the chain of steps an attacker would need to turn a weakness into real damage, for example moving from an exposed input to data they should not be able to read. A scanner might flag a suspicious function. Confirming the attack path shows whether that function can actually be reached and abused in the running application.

Turning AI agents loose on a live application raises an obvious question: what stops them from going too far? OpenAI's answer is containment. SecurityWeek reports that the agents work in isolated environments, with what OpenAI calls Guardian review enforcing testing boundaries. As described, Guardian limits which applications and endpoints the agents can reach during a test. The sources do not explain in technical detail how Guardian works.

Help Net Security and SecurityWeek both report that Red Team is available through an early-access program, not as a general release.

Defense Factory: from finding to pull request

Defense Factory covers the other half of the work. OpenAI's own product page describes it simply: "Agents trace vulnerabilities through your code and develop patches. Engineers review and test proposed patches before merge."

Teams feed it findings from their existing security tools and from Daybreak Red Team. Agents investigate each issue and develop a fix. According to OpenAI's product page, engineers can work with Codex to add context, examine evidence and tune a patch. SecurityWeek adds that the product supports importing vulnerability information from Wiz, Snyk and Endor Labs, and that engineers can create draft GitHub pull requests. Scans can run on new commits or on a schedule, so the tool can sit inside a normal development pipeline instead of running as a one-off audit.

Help Net Security adds detail on how it plugs in. Defense Factory, labeled as a preview, ships as a ChatGPT plugin. Teams connect a GitHub repository, and cloud scans run in Codex cloud. Each person who starts a scan needs a personal GitHub connection. Scans can cover a repository's default branch or only selected folders and files.

That last option matters in practice. A large monorepo may hold code that a security team does not own or does not want touched. Scoping scans to specific paths lets a team start with the code it is responsible for.

Where the human stays in the loop

The sources make the same point: Defense Factory does not merge its own code. Engineers review and test the proposed patches before they go in. The workflow OpenAI describes looks like this:

  • A code scanner, bug bounty submission, disclosure report or Red Team agent surfaces a possible vulnerability.
  • Red Team agents investigate it in an isolated environment and confirm whether it can be exploited.
  • Defense Factory agents trace the issue through the code and prepare a patch.
  • Red Team can test whether the patch blocks the exploit.
  • An engineer reviews, tests and decides whether to merge.

This is a real change in what AI security tools are being asked to do. Daybreak's agents act: they probe applications and propose changes to source code. The review step sits between those proposed changes and the merge, which makes the quality of that review the key control.

What It Means for Defenders

This is a product launch, not a vulnerability, so there is nothing to patch and no versions to update. But teams thinking about these tools, or about agentic security tooling in general, should plan for a few things.

Treat agent patches like any outside contribution. OpenAI says engineers review and test patches before merge. Make that rule enforceable, not just a habit: require normal code review and passing tests on agent-authored changes, and do not create a fast lane because a security tool wrote the code. A patch that closes one hole can still break something else or open a new one.

Use the retest step. According to Help Net Security, Red Team can check whether a patch actually blocks the exploit. If you adopt both tools, build that retest into your process so "fixed" means verified, not just merged.

Review the GitHub access model. Defense Factory connects to GitHub repositories, and each person starting a scan needs a personal GitHub connection. Security and platform teams should know who has connected accounts, which repositories those accounts can reach, and how access is removed when someone changes roles or leaves. Use the option to scope scans to selected folders or files where full-repository access is not needed.

Know where your code goes. Cloud scans run in Codex cloud, according to Help Net Security, so source code is processed outside your environment. Organizations with data-handling rules should check that before connecting a repository. That includes federal contractors working toward CMMC and anyone holding code tied to controlled information. This is a policy question for your compliance team, not something the sources settle.

Define testing boundaries before you test. OpenAI says Red Team agents run in isolated environments with Guardian review limiting what they can reach. Even so, decide which applications and endpoints are in scope before a test starts, get the right authorizations, and make sure nobody points an agent at systems you do not own. The usual rules of engagement for penetration testing still apply when the tester is software.

Expect access to be limited at first. Red Team is in early access and Defense Factory is in preview. Teams that want to evaluate them should plan a controlled pilot on a non-critical repository or application before relying on either tool for production work.

The bigger shift is clear even at this early stage. The work of confirming findings and drafting fixes, which has long been the slowest part of vulnerability management, is now something vendors are trying to hand to software. Teams that set strong review gates now will be in a better position to use that speed safely.

Sources