OpenAI Unveils Defense Factory to Automate Bug Patching

OpenAI has launched The Defense Factory, an open reference architecture designed for continuous, agent-driven vulnerability discovery and remediation. The release follows an internal security sprint where OpenAI mobilized more than 250 people across over 100 service areas. During this initiative, the company used its own frontier models to automate the security pipeline. The system closed 53 urgent or high-priority issues on its first day. Ultimately, agent-routed ownership assignments achieved a 90.6% acceptance rate, while 37% of findings were flagged as duplicates, and 19.5% of vulnerabilities were successfully reproduced at runtime.

The architecture relies on a continuous loop of inventory, discovery, dynamic validation, ownership assignment, and verified remediation, all coordinated via a shared SECURITY.md file. To execute these tasks, the system deploys agents inside ephemeral, isolated development containers. These agents leverage a mix of general-purpose models, including Sol, Terra, Luna, and a model OpenAI refers to as GPT-5.6, alongside specialized cyber models called Daybreak Blue for defensive scanning and Daybreak Red for triage and validation. Remediation in the sprint was 100% Codex-based, yielding a remarkably low 0.53% rolled-back fix rate and a 0.81% false-positive rate after dynamic validation.

Alongside the blueprint, OpenAI is releasing a Codex Security plugin, a playbook PDF, and an application for gated access to the Daybreak cyber models for authorized defensive work. The agents integrate with existing developer tools like GitHub, GitLab, Snyk, Semgrep, Tenable, Wiz, Jira, and Linear through model context protocols, command-line interfaces, or APIs.

For security practitioners, this release signals a shift away from traditional static analysis tools toward active, automated defense. Instead of managing a stagnant backlog of alerts, teams must now focus on building reproducible development environments and robust agent orchestration. Because attackers are already using open-weight models to accelerate exploitation, defenders must leverage their direct code access and frontier models to run security operations at agent speed.