OpenAI confirmed the launch of its new security line within Codex, in an announcement made through its official X account. The service, which is offered for now as a research preview, is intended to operate permanently in the cloud using agents: the analysis continues even when development teams shut down their computers. In this way, OpenAI positions Codex as a persistent defensive engineering assistant, not as a simple code scanning function.

The novelty is available to ChatGPT Pro, Business, Enterprise, and Edu users through a plugin on desktop and web Codex. Teams connect their GitHub repositories, select a supported cloud environment, and launch a full scan or continuous monitoring of commits. According to the company, the initial scan creates a project-specific threat model and examines the history, while subsequent scans focus on newly introduced code.

Unlike conventional static scanners, which compare code against predefined rules, Codex Security behaves like a security researcher: it reads the codebase as a whole, runs tests, examines realistic attack paths, and validates candidate vulnerabilities within an isolated environment before displaying them. Findings can include affected code, severity, validation evidence, guidance for remediation, and a proposed patch that the developer can inspect before creating a draft pull request.

By default, the service includes access to models with cyber capabilities using Daybreak Blue, which supports authorized defensive work such as vulnerability discovery, triage, secure code review, threat modeling, incident response, malware analysis, and patch validation; that access applies only within the Cloud product and does not extend to other Codex Security products or the API. For security teams, the greatest operational benefit is reduced alert fatigue: Codex investigates and removes duplicate results before submitting them, and scheduled assessments continue regardless of on-premises hardware.

OpenAI warns that the service requires governance: repository permissions must follow the principle of least privilege, cloud environments must restrict secrets and network access, and every patch generated must go through human review and testing before being merged. Documentation keeps humans on the approval path: reviewing the evidence, requesting a fix, examining the resulting patch, and deciding whether to open the pull request.

Sources: Cyber Security News, elhacker.NET blog , and OpenAI's Codex Security documentation.

Image credit: "Person holding the OpenAI icon" by FoxTPNL, via Wikimedia Commons, licensed under CC BY 4.0.