OpenAI moved Codex off the developer’s machine this week. Reusable cloud environments now hold a project’s repositories, dependencies and tools so that tasks start quickly and a team shares one approved setup, and each task gets its own workspace that keeps running while the laptop is asleep.
The documentation describes the setup flow plainly: pick GitHub repositories, let Codex inspect them, install dependencies and test the workflow, review the results, publish the environment, then start tasks against it. Internet access, environment variables and network secrets are configured per environment. Cloud Codex is available on Plus, Pro, Business, Healthcare, Education and Enterprise.
The security agent is the notable part
The release that will get security teams’ attention is Codex Security Cloud, a plugin that scans connected GitHub repositories in Codex cloud. It runs either a one-off repository scan or continuous monitoring of new commits, uses repository context to identify likely vulnerabilities, and validates issues in an isolated environment when it can. Codex investigates findings, removes duplicates and prepares fixes.
One line in the DevDay recap deserves attention. Codex Security Cloud, OpenAI says, “includes access to models offered through Daybreak Blue without a separate Daybreak application”. Daybreak is OpenAI’s gated programme for cyber-capable models. A product that ships with that access built in lowers the bar to models that were, until now, behind an application process.
The documentation is more cautious than the keynote. Codex Security Cloud is described as available in research preview on the web and in the desktop app, running scans requires Codex Security access, and for best results OpenAI recommends an account verified for Trusted Access for Cyber. The separate command line tool and TypeScript SDK ship as the public @openai/codex-security package and support bulk scans across repositories, runs in continuous integration, SARIF upload and a severity policy.
The rest of the developer release
The Codex command line interface was refreshed: tasks can be started and steered by voice, a new /agents view tracks several delegated tasks at once, and prompts, sessions and git worktrees are easier to handle. A code review experience in the ChatGPT desktop app reads summaries and diffs and answers questions about a change before feedback goes to a GitHub pull request or a GitLab merge request, with automatic reviews taking a first pass in the cloud.
What to watch
A scheduled scanner that prepares its own fixes is an agent with write access to a codebase, and the interesting failure mode is not a missed vulnerability but a confident wrong patch. OpenAI says findings are validated in an isolated environment “when possible”, which is the qualifier to watch. The second thing to follow is Daybreak: bundling cyber model access into a product is a change in how that access is governed, and OpenAI has not said whether the gating still applies to what the scanner can be asked to do.
