Codex now scans your repos while your laptop is closed

OpenAI's security agent grew up at DevDay. Codex Security Cloud is a research preview that scans connected GitHub repositories from OpenAI's cloud: on demand, on a schedule, or on every new commit. It builds a project-specific threat model first, then investigates findings, deduplicates them, validates them in an isolated environment, and proposes patches. "Fix with Codex" turns a finding into a draft pull request for a human to review. The laptop-closed line is OpenAI's own: scanning and investigation keep running in the cloud while your machine is off.

The workflow, as described in OpenAI's product docs via launch coverage, is more considered than "run the scanner harder." The threat model comes first: attacker entry points, trust boundaries, sensitive data, important code paths, modeled per project and editable by the team. Repository scans assess the full codebase and can review history for existing vulnerabilities. Commit scans watch new changes through monitoring you can enable, pause, and configure. Findings arrive with affected code, severity, validation evidence, and remediation guidance, not just a CVE string and a shrug.

Two details separate this from the static scanners it competes with. First, the validation step: findings are investigated in an isolated environment before they reach you, which is supposed to mean fewer phantom alerts. Second, the dedup: during an internal sprint, OpenAI says 36% of findings were identified as duplicates, and generated patches saw a 1% rollback rate. Those numbers come to us third-hand, OpenAI to a keynote to Simon Willison's live notes to a roundup, so treat them as the company's best-case anecdote, not a measurement. But the direction is right. Deduplication is where AI scanners earn their keep or drown you in noise.

The cyber-model angle is new. Codex Security Cloud includes access to cyber-capable models through Daybreak Blue by default, no separate Daybreak application required. That is OpenAI putting its offensive-security-grade models behind a defensive product, which is either reassuring or alarming depending on your priors, and it lands the same week OpenAI shelved a model for deceptive behavior. For the evaluation, what matters is whether the models find real vulnerabilities at an acceptable false-positive rate, not what they are called.

Availability: research preview for ChatGPT Pro, Business, Enterprise, and Edu users, through a plugin in Codex on desktop and web, part of the broader Codex push announced Tuesday. Connect a GitHub repo, pick a cloud environment, choose full scan or commit monitoring. Note the lineage: this extends the Codex Security research preview from March, which in its first 30 days scanned over 1.2 million commits and surfaced 792 critical and 10,561 high-severity findings, per an independent write-up. The product has six months of operational history. The cloud version is the new part.

The evaluation checklist, before this touches anything you care about:

1. False-positive rate on your codebase, not OpenAI's. Run it on a repo where you know the ground truth.
2. Patch quality. A draft PR is a proposal, not a fix. Read the diff. Run your tests. The 1% rollback figure is unaudited.
3. Permission scope. It needs GitHub access to read your code and write draft PRs. Scope it to non-production repos first and check exactly what the plugin can touch.
4. Data handling. Your code goes to OpenAI's cloud for analysis. Confirm retention terms against your own policies before connecting proprietary code.
5. Alert routing. Commit monitoring that nobody triages is just a more expensive way to ignore vulnerabilities. Decide who owns the findings queue before you enable it.

The honest framing: this is the shape application security is clearly taking, agentic investigation plus proposed remediation, from every vendor at once. OpenAI's version has the advantage of living where the code already gets written. Whether it is better than the alternatives is an empirical question your team can answer in an afternoon on a single repo. Do that before you form an opinion from a keynote.

What remains unknown: measured false-positive and true-positive rates on third-party codebases, the data retention and residency terms for scanned code, pricing beyond the research preview, and how it compares head-to-head with the incumbent scanners it wants to replace. The threat-model-first design is genuinely good product thinking. The proof is in the findings queue.