OpenAI Open-Sources Codex Security CLI for AI-Powered Code Vulnerability Scanning

OpenAI has released the Codex Security CLI as an open-source tool, giving developers and security teams a way to find, track and fix vulnerabilities in their code repositories using AI-driven analysis. The release, published to GitHub under the Apache 2.0 licence as `openai/codex-security` and available via the npm package `@openai/codex-security`, arrived so quietly that the Hacker News community discovered it before OpenAI had said a word publicly.

That’s not a bad way to build early credibility, as it happens. When a technically literate community finds your tool independently and starts poking at it, that’s informal peer review before the marketing machine even switches on.

What Codex Security CLI Actually Does

The tool is built on OpenAI’s Codex agent stack — the same underlying system that previously appeared as a research preview for ChatGPT Enterprise, Business and Edu customers, developed internally under the codename “Aardvark.” It entered research preview around March 2026 for certain enterprise and education users before this wider open-source release.

In practical terms, the CLI can scan entire repositories or selected files, review staged and unstaged changes before a commit goes in, compare findings across multiple scans, and export results in several formats: human-readable reports, JSON, SARIF and CSV. That last format matters because SARIF is the standard used by many existing security dashboards, which means Codex Security CLI can slot into toolchains teams already use rather than demanding they rebuild everything around it.

It also supports bulk scanning across multiple repositories at once, deduplication of findings between runs, false-positive tracking and historical result comparison. There’s even a `–max-cost` flag for budget controls, which will be welcome news for smaller teams watching their API spend.

Open Source — But Not Fully Local

Here’s where it gets a bit more complicated. What OpenAI has open-sourced is the client-side tooling: the CLI itself, the TypeScript SDK, bundled plugin workflows, Python helpers and the workbench used to manage scan state. The underlying AI models and analysis services remain hosted by OpenAI and aren’t running locally on your machine.

That means you’ll need an OpenAI account or API key to use it, and access is currently limited to approved customers and partners through a beta programme. The tool is clearly labelled as an early release.

Some commentators have been quick to flag that distinction. While the Apache 2.0 licence lets organisations inspect, modify and integrate the client code freely, the actual analysis depends on sending your code to OpenAI’s servers. For teams working in sensitive or regulated environments — financial services, healthcare, government — that raises real questions about data handling, intellectual property and vendor dependency.

The National Cyber Security Centre continues to advise that AI-based scanning tools should sit alongside established secure development practices, not replace them. Manual code review and traditional testing still matter.

What Developers Are Getting in Practice

For a developer on a typical engineering team, the workflow looks straightforward. Install via `npm install @openai/codex-security`, or just run it directly with `npx @openai/codex-security@latest –help`. You’ll need Node.js 22 or later and Python 3.10 or later, and it runs on macOS, Linux and Windows.

The CI/CD integration is probably where most teams will find the most value. The CLI can be wired into merge pipelines to enforce security checks before code lands in production — catching vulnerabilities at the point where they’re cheapest to fix, rather than after deployment.

Scan history is stored locally using SQLite, so teams can track how their security posture changes over time without relying on an external dashboard.

OpenAI hasn’t officially confirmed specific numbers around vulnerabilities fixed to date, so treat any figures circulating online with appropriate scepticism until the company publishes verified data.

What Happens Next

The beta status means the tool will evolve. OpenAI is presumably watching how approved customers use it before opening access more broadly. The open-source client also invites community contributions — bug fixes, additional workflow plugins, integrations with other tooling — which could accelerate development in ways a purely internal team couldn’t match.

Kevin Weil, OpenAI’s Chief Product Officer, has previously spoken about the company’s direction on developer tooling, though he hasn’t commented specifically on this release. The broader pattern, though, is consistent: open the client layer, keep the models proprietary, and let the developer community build the ecosystem.

Whether that model satisfies enterprise security teams who want full transparency over every component in their stack remains an open question.

What This Means for Kent Residents

Software developers, DevOps engineers and cybersecurity professionals working across Kent — including digital teams at Kent County Council, NHS Kent and Medway ICB, and the University of Kent — can download and use the open-source Codex Security CLI client as part of their security workflows, provided they have the right OpenAI access and meet the runtime requirements. For Kent-based tech SMEs and startups, the Apache 2.0 licence means the client code can be audited and adapted before committing to it, which lowers the barrier to trying it out. That said, any organisation in Kent considering it will need to think carefully about UK GDPR and sector-specific information governance rules before sending source code or configuration data to an external AI service — that applies here just as it would with any cloud-based analysis tool.