Security Health GitHub Action

The Security Recipes Health Check action turns this site’s guidance into a CI gate. It connects to the hosted Security Recipes MCP server for recipe context, evaluates bounded repository evidence with the model you choose, and reports every check — including the ones you have not enabled — in the job summary, so remaining coverage is always visible.

Quick start

name: Security health
on:
  pull_request:

jobs:
  security-health:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v5
      - uses: stevologic/security-recipes.ai/actions/security-health@main
        with:
          provider: openai              # anthropic | openai | xai | ollama
          api-key: ${{ secrets.OPENAI_API_KEY }}

Each check is a boolean toggle grounded in a published recipe:

Toggle	Default	Grounded in
`check-dependencies`	on	Vulnerable dependencies
`check-secrets`	on	Secrets and data exposure audit
`check-injection`	on	Injection sink audit
`check-supply-chain`	on	Supply chain build integrity audit
`check-authz`	off	Authorization and tenant boundary audit
`check-containers`	off	Base image hygiene
`check-owasp`	off	OWASP Top 10 audit
`check-cve-exposure`	off	CVE intelligence intake gate
`check-compliance`	off	Compliance standards

Models

Select provider (Anthropic, OpenAI, Grok/xAI, or an Ollama-served model via base-url) and optionally model. When no model is set, the action queries the provider’s live model list and picks the lowest available model — for Ollama, the smallest installed model by size — so the default run stays cheap.

How verdicts work

Recipe context is fetched from https://security-recipes.ai/mcp (streamable-HTTP MCP) with an automatic fallback to the public api/recipes.json feed.
Evidence is bounded and deterministic: manifests, workflows, container files, and source excerpts, capped to a few dozen kilobytes per check.
The model must answer in strict JSON (pass / warn / fail plus findings); malformed replies degrade to warn, never to silence.
fail-on decides whether the job fails on fail (default), on warn, or never (report-only).

This repository runs the action on itself (workflow) in report-only mode: the deterministic build gate stays the required check, and the LLM verdicts surface loudly in every pull request summary.

An LLM check is a bounded review gate, not a proof of security. Keep the evaluation discipline and your required human review and CI gates in place.