OpenAI has published the tooling behind its AI-driven vulnerability hunting as open source. The repository openai/codex-security, released alongside the company's Daybreak cyber program, ships a command-line tool and a TypeScript SDK — published to npm as @openai/codex-security — that scan repositories, selected paths or Git diffs for security flaws.
According to the project's documentation, Codex Security does more than flag suspicious code. It "validates candidate findings, generates patches, and verifies existing fixes"; it can draft SECURITY.md policies and save threat models; and it can browse saved scans and findings, identify duplicates, assess severity against a user-defined rubric and suggest owners drawn from source and Git history. Deep scans run parallel discovery workers across a repository or chosen paths.
The tool is designed to fit three modes of use. Inside a Codex session, a developer enables the Codex Security plugin and asks it to run a scan; from a terminal, installing the CLI globally (with `cs` as a short alias) and running `cs scan .` produces a saved report, findings, coverage and, where available, a threat model; in automation, the TypeScript SDK or a supplied GitHub Action can run scheduled, manual or pull-request scans, with results exportable as SARIF, JSON or CSV and publishable to services such as Linear. The documentation notes a minimum of Node.js 22.13 within the 22.x line, or Node.js 24.x/26.x, plus Python 3.10 or newer.
Notably, the scanner is not locked to OpenAI's own models: it supports Amazon Bedrock, OpenRouter, Fireworks AI and Cheaper Inference as alternative providers. Findings are report-only by default, with an optional severity threshold that can be set to fail a CI job.
The release lands after OpenAI disclosed that AI-assisted scanning had helped surface hundreds of previously unknown flaws — including 948 Linux kernel privilege-escalation exploits and 81 unknown Firefox vulnerabilities reported in earlier coverage. Making the workflow available as open source lowers the barrier for teams to run the same kind of automated security review on their own code, and puts a concrete, scriptable tool behind the company's broader pitch that models can meaningfully assist defensive security work.
For teams evaluating it, the practical questions are mundane but important: how well the deep scans handle a large monorepo, what the false-positive rate looks like against a project's own rubric, and how the tool behaves when the required sandbox tooling is unavailable. Those are the kind of details the open repository now lets engineers test directly rather than take on faith.

