Best for security engineers and small teams who want a validated first pass over a repository and its new commits without running scanners by hand. It supplements static analysis rather than replacing it, and a person still approves every patch. Expect a setup step for GitHub and a Codex cloud environment.
OpenAI launched Codex Security on September 29, 2026, an agent that hunts vulnerabilities in code. The cloud version reads GitHub repositories once or on a recurring schedule, watches new commits, reproduces likely flaws in an isolated container, and drafts fixes for review. It is a research preview for Pro, Business, Enterprise, and Edu users.
Maker: OpenAI · Autonomy: semi autonomous · Maturity: BETA
Underlying models: Models offered through OpenAI Daybreak Blue
About Codex Security
OpenAI introduced Codex Security at DevDay on September 29, 2026, next to the wider OpenAI Codex updates. OpenAI describes it as a tool that helps security and engineering teams find, confirm, and fix vulnerabilities. The cloud version reads whole GitHub repositories when asked or at set intervals, keeps checking new commits, investigates what it finds, removes duplicates, and prepares fixes in the cloud even with your laptop closed. That plan-act-verify loop is what makes it an agent rather than a code linter that prints a list.
Each scan follows four stages in OpenAI's documentation. Codex builds a threat model of the repository, scans it once or watches commit changes, tries to reproduce likely vulnerabilities in an isolated container to cut false positives, and then proposes a minimal patch. Findings arrive ranked by criticality with the validation evidence attached. Nothing is applied automatically: a proposed patch becomes a draft pull request only after you select it. OpenAI says the tool complements static analysis rather than replacing it, and it does not replace human review.
There are two ways to run it. Codex Security Cloud is a plugin for ChatGPT on the web and in the desktop app, currently a research preview, that needs a connected GitHub repository and a Codex cloud environment. A separate local plugin, plus a public CLI and TypeScript SDK published as @openai/codex-security, covers terminal scans, CI checks on pull requests, and bulk scans across many repositories. The cloud version includes access to models offered through Daybreak Blue, so you do not file a separate Daybreak application. That differs from GPT-5.6-Cyber, which sits in the separate Daybreak Red tier for vetted partners.
Availability covers Pro, Business, Enterprise, and Edu plans on desktop and web. OpenAI did not publish a separate Codex Security price at review time, so it appears to ride on the plan you already pay for. It suits AppSec engineers and small teams that want a validated first pass over a codebase. Teams that mainly need a general coding assistant should compare GitHub Copilot, Claude Code, and Jules, and OpenAI's own dots and ChatGPT Work cover different jobs. The vendor behind all of these is OpenAI.
Pricing
OpenAI lists no separate price for Codex Security at the time of review. Access to Codex Security Cloud is included with four ChatGPT plan types (see the vendor announcement), and it runs scans in Codex cloud, so it needs a workspace with Codex cloud set up. Command-line and SDK users can cap estimated spend for a scan. Running scans requires Codex Security access, and OpenAI recommends an account verified for Trusted Access for Cyber for best results. Check the plan page in ChatGPT for current plan prices.
Key Features
- Repository and commit scans: Scans a whole GitHub repository once, or monitors new commits and can review existing commit history, with monitoring pausable per repository.
- Threat model per repository: Builds an editable threat model of entry points, trust boundaries, and risky components that steers what gets scanned and how findings are ranked.
- Sandbox validation: Tries to reproduce each likely vulnerability in a clean isolated container and marks the ones that reproduce as validated, attaching logs and commands as evidence.
- Proposed patches: Offers Fix with Codex on findings that support it, producing a minimal diff you review before creating a draft pull request.
- CLI, CI, and SDK: The @openai/codex-security package runs local and bulk scans, checks pull requests in CI, uploads SARIF, and exposes a TypeScript SDK.
- Daybreak Blue models included: The cloud plugin bundles Daybreak Blue model access, so defenders skip the separate Daybreak application step.
Strengths
- Tries to reproduce each suspected issue in a clean container first, so validated findings arrive with logs and commands rather than a bare warning.
- Runs in Codex cloud on a schedule or on new commits, so scans continue while your laptop is closed.
- Never applies a patch on its own: a proposed fix waits behind a draft pull request that a person has to create.
- Ships as a cloud plugin, a local plugin, a CLI, and a TypeScript SDK, so one scanner covers desktop, terminal, and CI use.
Weaknesses
- The cloud plugin is a research preview and needs a Codex cloud environment plus a GitHub connection, so setup is longer than a single-click install.
- OpenAI says it complements static analysis and does not replace manual review, so you still need existing scanners and a human in the loop.
- Scan time varies with repository size and validation work, and OpenAI gives no fixed duration or per-scan cost.
- Running scans needs Codex Security access, and OpenAI recommends a Trusted Access for Cyber verified account for best results.
Frequently Asked Questions
Does Codex Security have its own price?
OpenAI's documentation and DevDay recap list no separate price for Codex Security. Codex Security Cloud is offered to ChatGPT Pro, Business, Enterprise, and Edu users, so it appears to be covered by the plan you already hold. The only spending control OpenAI describes is an estimated cost limit you can set when scanning from the CLI or SDK.
Can Codex Security replace our existing scanners?
According to OpenAI it sits alongside static analysis instead of taking its place, because deterministic scanners still give broad coverage. It adds model-based reasoning and an attempt to reproduce each finding in a sandbox. It also states that it does not replace manual security review or human threat assessment.
What else could you use for security-minded code work?
Claude Code and Jules are general coding agents that write and change code, so they suit fixing what a scan reports rather than producing the scan itself. GitHub Copilot fits teams already standardised on GitHub tooling for everyday coding. None of the three is described here as a validated vulnerability scanner, so check each vendor's security documentation before relying on it.
How does Codex Security relate to OpenAI Codex?
Codex Security runs on the Codex platform: the cloud version runs scans in Codex cloud, and the local plugin runs each scan as a Codex task in the desktop app or CLI. Codex is the general coding agent, while Codex Security is a purpose-built scanner with a threat model, validation stage, and findings list.
How do you run a first scan?
Open Plugins in ChatGPT on the web or desktop, install Codex Security Cloud, and select New scan. Connect GitHub if prompted, pick the repository and a Codex cloud environment, choose Repository under What to scan, then start the scan. For a local scan, run npx @openai/codex-security --help and follow the CLI quickstart.
Top Alternatives
- OpenAI Codex: Pick Codex for writing and reviewing everyday code changes; pick Codex Security when the job is finding and validating vulnerabilities.
- Claude Code: Pick Claude Code as a general terminal coding agent; pick Codex Security for scheduled repository scanning with sandbox validation.
- Jules: Pick Jules for asynchronous coding tasks on your repos; pick Codex Security for a ranked, validated list of security issues.