Skip to content
D-CSIL™

AI News · 2026-11-13 · 1:04 PM CT

OpenAI's agents attack your app, then draft the fix

Paul Derrington, Jr.

By Paul Derrington, Jr.

CEO | Chief Engineer | Research Engineer | Editor, D-CSIL

3 min read

Follow @PaulDerrington on X

Key Takeaways

OpenAI added two Daybreak products on November 12: Red Team, in early access, uses agents to pen-test web apps and confirm which reported flaws are real, and Defense Factory, in preview, traces confirmed flaws and prepares patches that engineers must review before merging. OpenAI did not give pricing for either, and no independent test results were found at publication.

A loop, not a scanner

OpenAI expanded its Daybreak cybersecurity program on November 12 at its Cyber at the Frontier event, adding two products that work as a pair. Daybreak Red Team, in early access, sends coordinated AI agents at a web application to find and confirm vulnerabilities. Daybreak Defense Factory, in preview, takes confirmed findings and prepares the fix.

This is not a new model launch. The models underneath Daybreak — OpenAI lists general-purpose models plus its Daybreak Blue and Daybreak Red security models — were already gated behind Daybreak applications. What is new is the workflow around them: testing, validation, patching and retesting, offered as products. SecurityWeek reports that OpenAI is accepting early-access requests for Red Team and that Defense Factory is offered through a preview plugin in ChatGPT for workspaces with Daybreak access.

From finding to fix

OpenAI says Red Team's agents use source code and application context to investigate attack paths and confirm vulnerabilities. The Deep View, reporting on OpenAI's press briefing, says one agent coordinates each assessment, others investigate the source code and application context, and reviewer agents check the results. It can do the same validation for findings from code scanners, bug-bounty submissions and disclosure reports, per OpenAI and Help Net Security's November 13 coverage.

Once a flaw is confirmed, Defense Factory's agents trace it through the codebase and prepare a proposed patch. An engineer reviews and tests that patch before anything merges, and can use Codex to examine the evidence, add context and refine the change. OpenAI says Red Team can then test whether the patch stops the exploit and repeat the assessment as the app changes. Defense Factory also ingests findings from existing security tools — Help Net Security notes teams install the Defense Factory (Preview) plugin in ChatGPT, connect a GitHub repository, and run cloud scans in Codex cloud, with each person starting a scan using a personal GitHub connection. SecurityWeek reports that Defense Factory can import vulnerability information from Wiz, Snyk and Endor Labs, and that engineers can create draft GitHub pull requests.

The guardrails are part of the product

Access to both products runs through Daybreak. OpenAI says Daybreak applicants are verified and their planned use is reviewed before access is approved. According to OpenAI, Red Team assessments run inside isolated environments, with Guardian review enforcing the assessment scope, and customers control which applications and endpoints the agents can reach.

On the fix side, OpenAI says engineers review and test proposed patches before merge, and SecurityWeek reports that patches are not deployed automatically. Inside its own Defense Factory operation, OpenAI describes the same pattern — a five-step loop of inventory, discovery, dynamic validation, ownership assignment and verified remediation, with people reviewing consequential changes and independently verifying deployed fixes.

What has not been shown yet

OpenAI did not give pricing for either product, The Deep View reports, and no independent evaluation of Red Team's detection rate, false positives or patch quality was found for this briefing. The Register reported Nancy Wang, who leads cyber-applied AI at OpenAI, describing Red Team as “superhuman pen-test at the scale of automated continuous scanning.” OpenAI's Red Team product page does not publish detection or accuracy figures.

The hardest numbers OpenAI does publish come from its own internal security sprint, not from customer use of these two products: it says it mobilized more than 250 people, found 37% of findings were duplicates, reproduced 19.5% of findings at runtime, and got the false-positive rate after dynamic validation down to 0.81%. OpenAI also reports a 0.53% rolled-back fix rate from that work.

Disclaimer: Derrington Collaborative AI is my personal business, operated independently of and outside my official U.S. Government position and duties. The views and opinions expressed on this site and on any and all of my social media accounts are my own. They do not reflect the official policy or position of the Department of the Army, the Department of Defense, or the U.S. Government. Mention of any company, product, or service does not constitute an official endorsement.