OpenAI has published Codex Security, a command-line tool and TypeScript SDK that scans a repository for vulnerabilities, validates the findings, drafts patches, and verifies that existing fixes actually hold. The CLI installs from npm as @openai/codex-security, signs in with cs login, and runs with cs scan .. It is not a demo. It ships with a GitHub Action, a Docker Compose setup for bulk scans, a policy command that drafts a SECURITY.md, and export to SARIF, JSON and CSV.

The interesting part is not that OpenAI built a security agent. Plenty of vendors have bolted a scanner onto a model. The interesting part is the provider list. The docs say scans support OpenAI, Amazon Bedrock, OpenRouter, Fireworks AI and a service called Cheaper Inference. Bedrock uses AWS credentials and “does not require a separate OpenAI login.” That is a strange sentence to find in an OpenAI repository. The company is selling a security workflow that, by design, does not need OpenAI’s models to run.

What the tool actually does

The scan surface is broader than a grep for known-bad patterns. You can scan a whole repo, selected paths (--path src --path tests), or a diff against a base revision (--diff origin/main). A --mode deep scan repeats Standard scans and returns deduplicated findings, which tells you the team expects the tool to over-report and has built machinery to collapse the noise. There is an advisory time budget, --max-time 15, for a fifteen-minute Standard investigation.

Beyond detection, the workflow is agentic. A completed scan saves a report and findings with source evidence and exclusions. From there you can ask Codex to validate a candidate finding, generate a patch, or verify a fix that already landed. It will also suggest owners “from source and Git history,” assess severity against your own rubric, and track findings in Linear, Jira or GitHub. The policy command drafts a SECURITY.md outside the repository and leaves existing files untouched, which is the right default for a tool that might otherwise clobber a hand-written disclosure policy.

The plumbing is ordinary and that is the point. Node.js 22.13.0+ within 22.x, or 24.x or 26.x, plus Python 3.10+. For CI you set OPENAI_API_KEY or CODEX_API_KEY. The GitHub Action example uses an estimated $5 limit per scan and reports findings without failing the job. A weekly scan on GitHub-hosted Ubuntu. This is a tool built to be dropped into an existing pipeline, not a new platform to migrate onto.

The provider list is the story

Read the “Other providers” section twice. Amazon Bedrock, OpenRouter, Fireworks AI, Cheaper Inference. Those are competitors to OpenAI’s API business, and Codex Security treats them as first-class backends. You set the provider’s key, choose a supported model, and scan.

There are a few ways to read this and only one of them is generous. The generous read: security scanning is a workflow product, and the workflow is the product. OpenAI would rather own the agent loop, the report format, the deduplication, the SARIF export and the Jira integration than fight a provider war inside a cs scan command. The cynical read: the scanner is a loss leader for the Codex ecosystem, and letting it run on Bedrock costs OpenAI little because the real lock-in is CODEX_SECURITY_STATE_DIR and the Cloud findings service, not the token bill.

Both can be true. What is not in dispute is the admission buried in that section. If a security agent’s value were mostly the model, you would not ship a config path for three rival inference providers on day one. The value is in the harness: the scan scope logic, the deduplication, the validation loop, the owner suggestion from Git blame, the export formats that ServiceNow and CI systems already ingest.

Where the model still matters

The tool is candid about its own limits in a way that is worth noting. Findings carry “coverage limits” and “exclusions” that the report is supposed to explain. The --max-time flag is advisory, and the docs warn that “requests in progress can exceed the limit.” A GitHub Action that reports findings without failing the job is a scanner that expects to be wrong sometimes and does not want to break your build over it.

Then there is the access tiering. “Some cybersecurity requests and protected findings require Trusted Access for Cyber,” and accounts with “Daybreak access” get a separate opt-in for the Codex toggle and “Cyber access programs.” OpenAI is gating the sharpest security capabilities behind an approval process. That is a policy decision dressed as a product feature, and it is the right instinct for a tool that can read your whole codebase and write patches to it. It also means the scanner most teams can run today is not the scanner OpenAI can run internally.

There is a smaller tell in the changelog voice of the docs: “The local serve command and browser dashboard have been removed; existing databases and scan artifacts remain available.” Someone shipped a dashboard, then pulled it. The tool is being iterated in public and the surface is still moving.

What it means for builders

The practical takeaway for anyone running a codebase: a $5-per-scan weekly job that produces SARIF and opens Jira tickets is now a realistic line item, and the provider flexibility means you are not forced onto one vendor’s pricing to get it. The strategic takeaway is less comfortable for OpenAI’s competitors and more interesting for everyone else. The company that makes the best model has decided that in security, the model is a component you can swap. If that logic holds for vulnerability scanning, it will be tested next on code review, incident response and the rest of the agentic devtools stack.

Watch whether Codex Security Cloud becomes the default sink for findings, and whether the Bedrock path survives the next release. A provider-agnostic scanner is a strong opening move. A provider-agnostic scanner that quietly stops supporting rivals six months in would tell you the flexibility was a wedge, not a principle.