Codex Security

• Running a first standard scan of a service before a release or an external security review
• Adding a CI job that scans each pull request's changes, uploads SARIF to GitHub code scanning, and fails on high-severity findings
• Scanning every non-archived repository in a GitHub organization as a resumable bulk campaign
• Triaging a backlog of SAST, Dependabot, or bug-bounty findings to see which ones affect the code you actually run
• Patching confirmed high and critical findings and opening a pull request with the verified fixes
• Monitoring new commits on a critical GitHub repository with Codex Security Cloud

Codex Security's strongest idea is honesty about coverage. Every scan records what it reviewed and what it skipped, and OpenAI's own docs warn that a missing finding alone does not prove a fix worked, pointing users to reruns, direct rechecks, and a verify-fix check instead. That caution suits real security programs better than a confident list of findings. The range of surfaces is also useful, since the same scanner covers an ad-hoc desktop scan, a CI gate, a PR reviewer, and continuous cloud monitoring. The weak points are access and cost. Availability is gated and described inconsistently across OpenAI's pages, there is no per-scan price, and AI scans can vary between runs. Run it next to your existing scanners on a scoped service first, and check the coverage file before trusting a clean result.

About Codex Security

Codex Security is OpenAI's dedicated application security product. It runs inside Codex as its own plugins, CLI, and pull-request reviewer, but instead of writing features it reviews code the way a security researcher would. A standard scan moves through threat modeling, discovery, validation, attack-path analysis, and reporting. A deep scan runs several workers for broader, longer review of a critical service or directory. Each finding records severity, confidence, location, evidence, and remediation guidance, and a separate coverage file lists what was reviewed, excluded, or deferred, so a scan with partial coverage is not mistaken for a clean result. Fixes are bounded: you choose which findings to patch, Codex Security prepares a focused change, and where it can, it adds a regression test that fails before the fix and passes after it. It does not apply patches on its own. There are four ways to use it. The Codex Security plugin adds a Security workbench (Scans, Findings, Repositories) to the ChatGPT desktop app and also works in the Codex CLI. The `@openai/codex-security` package, published with the plugin's workflows under Apache-2.0 in the openai/codex-security repository, provides a CLI and TypeScript SDK for local scans, bulk scans across a GitHub organization, a pre-commit hook, and CI jobs that upload SARIF and fail above a severity threshold. It can run with a ChatGPT sign-in, an OpenAI API key, or models from Amazon Bedrock, OpenRouter, or Fireworks. Codex Security Review adds security-focused reviews to GitHub pull requests. Codex Security Cloud, a separate plugin that OpenAI's docs describe as a research preview, scans connected GitHub repositories in ephemeral Codex cloud containers, either once or continuously as new commits land. Access and cost need the most attention. The npm package is public, but running scans requires Codex Security access, and some full-repository scans may also need Trusted Access for Cyber verification. Pull-request reviews consume Codex allowance or ChatGPT credits, API-key scans are billed per token, and OpenAI publishes no per-scan price. OpenAI positions it as a complement to deterministic SAST tools, not a replacement, and says results can vary between runs of the same scan.