Codex Security CLI

What it is

@openai/codex-security — OpenAI's CLI/TypeScript SDK for defining security policy and finding, validating, and fixing vulnerabilities. Beyond scanning, it generates SECURITY.md policy drafts, runs a standalone findings service, deduplicates findings across repositories, and classifies severity against a custom rubric. Some requests and "protected findings" are gated behind OpenAI's Trusted Access for Cyber program (join at chatgpt.com/cyber).

Preconditions / Affected

• Node.js 22.13.0+ and Python 3.10+.
• Interactive use requires codex-security login; CI uses OPENAI_API_KEY instead.
• Certain cybersecurity requests require Trusted Access approval — a gate the summary omits.

What to do

Install and run a basic scan:

npm install @openai/codex-security
codex-security login
codex-security scan /path/to/directory

Generate policy guidance (repository-wide or component-scoped):

codex-security policy .
codex-security policy . --path services/api --knowledge-base architecture.md

Note the ordering constraint: policy saves a draft outside the checkout — it does not install it or run a scan. Review the diff before copying. Supporting threat-model/architecture docs stay outside the repo and may hold sensitive detail.

Publish and dedupe against a findings service:

codex-security publish scan --to custom --findings-url http://localhost:3000

dedupe retrieves candidates, runs independent Codex reviews locally, and persists accepted groups; --all-repositories widens scope.

Classify severity before filing tickets:

codex-security classify-severity --scan SCAN_ID --rubric /path/to/policy.md

This checkpoints each finding in SQLite and reuses assessments on rerun; --reprocess forces reassessment.

Alternate inference providers (verbatim examples):

codex-security scan . --provider amazon-bedrock --model openai.gpt-5.6-luna
codex-security scan . --provider openrouter --model anthropic/claude-sonnet-4.5
codex-security scan . --provider fireworks --model accounts/fireworks/models/qwen3-235b-a22b

For many repos, use the Docker Compose config. Run the findings service without Docker via codex-security serve (see "running without Docker" for credentials/storage). The service and scanner share the ghcr.io/openai/codex-security image; the findings service uses a separate container/volume via compose.findings.yaml.

Caveats

• classify-severity leaves the original scan severity unchanged; it only produces a separate rubric-based assessment.
• The /dashboard is read-only and refreshes every five seconds.
• Duplicate detection is embedding-similarity within a repo unless --all-repositories is set, and accepted duplicates still require an independent local Codex review.