AI Vulnerability Remediation Quick Start

Use this page when you want to try security-recipes.ai without adopting new tooling. Pick one real finding, give the agent one recipe, and keep the output reviewable.

The first win is not full automation. The first win is a clean PR or a useful triage note that followed a recipe and respected your repo rules.

The loop

  1. Pick one finding. Use a small dependency, SAST, secret, container, SDE, or CVE finding. Avoid a noisy backlog for the first run.
  2. Pick the matching recipe. Start in AI vulnerability remediation playbooks or search the Recipes.
  3. Choose the agent your team already uses. GitHub Copilot, Claude, Cursor, Codex, Devin, Shiba Studio, Hermes, and OpenClaw all work with this pattern.
  4. Put the rules where the agent reads them. Use the native instruction file for your agent.
  5. Add only the context needed. Attach scanner output, advisory details, SBOM/SARIF evidence, or an approved read-only MCP connector.
  6. Review the result. Accept a PR only when tests and reviewer expectations are satisfied. Otherwise keep the triage note.

Minimal agent task

Remediate one security finding using the matching security-recipes.ai recipe.

Finding:
- 

Rules:
- Use one recipe and cite it in the output.
- Make the smallest safe change.
- Do not touch unrelated files.
- Use MCP servers as read-only context unless this task explicitly grants write access.
- Run the relevant tests.
- Open one PR, or stop with a triage note if the fix is not bounded.

Where to put recipe rules

Agent File or place to start
GitHub Copilot .github/copilot-instructions.md plus a narrow issue assigned to @copilot
Claude CLAUDE.md; use a skill for repeatable remediation procedures
Cursor .cursor/rules/security-remediation.mdc
Codex AGENTS.md
Devin Knowledge entry or playbook attached to the task
Shiba Studio Per-agent instructions and skills on a workspace-bound agent
Hermes A dedicated remediation skill with the provider and sandbox backend pinned
OpenClaw Workspace AGENTS.md operating rules (persona boundaries stay in SOUL.md)

Starter repo rules

Copy this into the relevant agent instruction file and edit the commands to match your repo:

# Security remediation rules

- Work on one security finding at a time.
- Prefer the matching recipe from security-recipes.ai.
- Make the smallest safe change that remediates or clearly triages the finding.
- Do not edit migrations, production infrastructure, generated files, or release
  automation unless the task explicitly allows it.
- Run the tests for the touched area before proposing a PR.
- Include the finding ID, recipe used, files changed, tests run, and residual
  risk in the output.
- Stop with a triage note if ownership, blast radius, or verification is unclear.

Add MCP context only when it helps

If you use the Security Recipes MCP server, start with recipes_search or recipes_match_finding, then recipes_get or recipes_cve_get. Call recipes_quality_report only when you are choosing which recipe is ready enough to automate. Rechecked August 23, 2026: MCP 2026-07-28 is still current and stateless. Servers MUST implement server/discover. The optional server in this repository stays read-only.

Good first connectors are read-only:

  • GitHub or GitLab repository context.
  • Code scanning, Dependabot, SARIF, or SBOM data.
  • Public advisory/package data such as OSV, GitHub Advisories, deps.dev, package registries, or NVD-backed mirrors.
  • Product security tools your organization already approves, such as Semgrep or Snyk.

Do not start with write-capable tools. Ticket creation, code mutation, cloud changes, secret rotation, and deployment should remain outside the first run.

Good first recipes

Vulnerable DependenciesThe cleanest first loop for dependency and package advisory findings.SAST FindingsUse the recipe when the fix is local, testable, and tied to one rule.Sensitive DataUse for bounded cleanup of accidental secrets, PII, or sensitive fields.Base ImagesUse when the remediation is an image, Dockerfile, or OS package bump.

Next steps