Security

An AI Security Reviewer on Every PR: Codex Catches 3x What Semgrep Does.

OpenAI has released Codex Security Review. It analyzes GitHub PRs for vulnerabilities using full-repository context, then validates findings in a sandbox so only truly exploitable bugs are reported.

· Huuush

An AI Security Reviewer on Every PR: Codex Catches 3x What Semgrep Does

Key takeaways

  1. 1OpenAI has released Codex Security Review as a research preview. It analyzes GitHub pull requests for vulnerabilities using repository context and verifies whether they are exploitable.
  2. 2It works in three stages (identification, validation, remediation), reproduces vulnerabilities in an isolated sandbox, and suggests a minimal patch without applying it automatically.
  3. 3In an independent test it caught 74% of real bugs, ahead of Semgrep (20%) and Snyk (28%), but it works only with GitHub and the Plus plan is not supported.
Codex Security Review: automated security review for GitHub PRs
Codex Security Review finds and validates PR vulnerabilities using repository context (AI-generated image)

OpenAI has released Codex Security Review as a research preview. It automatically analyzes GitHub pull requests (PRs) for security vulnerabilities and reports only the bugs it has validated as actually exploitable.

Loading X post.

Unable to load the X post. Please try again.

View original on X
OpenAI Developers: official announcement of Codex Security Review (X embed)

Reading the PR and Proving the Vulnerability: How Codex Security Review Works

Existing security tools look only at the diff and find vulnerabilities by pattern matching. Codex uses the full repository context and a project-specific threat model. It maps attacker entry points, trust boundaries, sensitive data, and high-impact code paths to explore real attack paths.

Three Stages That Keep Only Real Vulnerabilities

1. Identification: analyze the repository, build a threat model, and explore attack paths
2. Validation: reproduce the issue in an isolated sandbox to confirm it is exploitable
3. Remediation: suggest a minimal patch that fixes the root cause (not applied automatically; a human reviews it)

Before You Add It to GitHub PRs: Setup and Scope

Comment @codex security review on a PR and it runs immediately. You can also set it to run automatically (when a PR is opened, or on every push). Findings appear as inline comments in the PR.

3x Semgrep: What More Did It Find?

In an independent test, Codex Security Review hit a 74% rate on real bugs. On the same codebase, Semgrep scored 20% and Snyk 28%. Traditional SAST tools have false positive rates of 30% to 70%, while Codex claims to cut false positives by more than 50% through context-based reasoning.

Real-World Results: CVEs Already Found

OpenAI says an internal version found real CVEs in well-known open source projects including GnuPG, GnuTLS, OpenSSH, PHP, and Chromium, and disclosed them responsibly. Scans turned up 792 critical and 10,561 high-severity findings.

Who Can Use It Now, and What Limits Remain

  • Available to: ChatGPT Enterprise, Edu, Business, and Pro, free during the introductory period
  • Not supported: the Plus plan
  • Limits: GitHub only, slow first scans on large repositories, and no automatic code fixes

When PR Review Becomes the Defense Line Before Deployment

This is a turning point from pattern matching to AI that understands and validates. The era of developers checking security by hand on every PR is shifting to one where AI validates first and humans review the results.

Related: OpenAI Help · Cybersecurity News · X post

Frequently asked questions

How do I run Codex Security Review?
Comment "@codex security review" on a PR and it runs immediately. You can also set it to run automatically when a PR is opened or on every push. Findings appear as inline comments in the PR.
How is it different from existing security tools?
Existing tools look only at the diff and find vulnerabilities by pattern matching. Codex uses the full repository context and a project-specific threat model to explore real attack paths, then reproduces them in an isolated sandbox to validate them.
How well does it perform?
In an independent test it caught 74% of real bugs, versus 20% for Semgrep and 28% for Snyk on the same codebase. It claims to cut false positives by more than 50% through context-based reasoning.
Who can use it?
It is free during the introductory period for ChatGPT Enterprise, Edu, Business, and Pro. The Plus plan is not supported. It works only with GitHub and does not fix code automatically.

Related articles