At its DevDay 2026 developer event on September 29, OpenAI unveiled Codex Security Cloud, the cloud edition of its Codex Security agent, which hunts for code vulnerabilities and drafts fixes. It reviews an entire connected GitHub repository and then keeps watching every commit added afterward. It is currently a research preview, available only on higher-tier ChatGPT plans.
What it does
Codex Security Cloud ships as a plugin for the Codex desktop and web apps. The scanning itself runs in Codex cloud environments, so work continues even when your laptop is closed and you can do something else while a large repository is being examined.
The process has three stages: spotting suspected vulnerabilities, investigating whether each suspicion is a real problem, and preparing a fix proposal. When the same issue appears in several places, the findings are deduplicated. Proposals are not merged automatically; a human reviews them before deciding whether to accept a change.
Looking at the whole repository, not single files
Earlier lightweight checks often operate on a changed file or a single pull request. Because Codex Security Cloud treats the whole repository as context, it can reportedly flag problems that span several elements, such as authentication paths, dependencies, configuration files and data flows.
Where possible, it tries to reproduce a suspected issue in an isolated environment to confirm that it looks exploitable before reporting it. That approach differs from tools that simply list warnings based on pattern matching.
Eligible plans and requirements
The feature is available on ChatGPT Pro, Business, Enterprise and Edu plans. Daybreak Blue, a cyber-capable model, is included by default, and no separate approval is said to be needed.
Only GitHub is supported. Teams on GitLab or other hosts are pointed to the command-line version of Codex Security instead. The CLI uses the package @openai/codex-security and supports CI integration and SARIF output.
Pricing and caveats
Billing is based on token usage, and no per-scan estimate has been published. Coverage has pointed out that scans pause when funding runs out, so it is safer to set a budget cap before trying a large repository. As a research preview, it carries no service-level guarantee or promise of general availability.
No concrete figures on detection rate, false-positive rate or fix quality have been shared yet. Because a model reasons its way to findings, identical code will not necessarily produce identical results on every run, so using it as a second pair of eyes alongside existing static analysis tools is more realistic than treating it as a replacement.
Summary
Codex Security Cloud is appealing because it hands whole-repository review and continuous commit monitoring to the cloud, but it is still a research preview with limited information on cost and accuracy. A sensible start is to pick two low-stakes repositories, set a spending cap, and compare its results with your current scanner.
