AI Development Methodology

Security Policy

What this repository is

This is a documentation and methodology project — markdown plus git, nothing else.

• No executable code. No JavaScript, Python, or other source to run.
• No package dependencies. No package.json, requirements.txt, lockfiles, or third-party packages in the delivered artifact.
• No adopter build step or bundled install scripts. The skill contains Markdown and supplies no executable components. The optional external skills installer runs software and is not part of this repository. There are no bundled preinstall / postinstall / prepare hooks because there is no package manifest.
• No telemetry, no network calls, no credential collection. The repo does not phone home, read your environment, or touch your secrets.
• One checked-in workflow. A secret scan (workflow) runs with read-only contents permission. It uses actions/checkout and gitleaks/gitleaks-action, pinned to full commit SHAs, and passes the repository GITHUB_TOKEN to the second action. GitHub also hosts the existing Pages build; _config.yml supplies static site metadata. These hosting/build services and the optional external installer are separate from the Markdown skill an adopter installs.

The delivered skill has no executable components or package dependencies. That is narrower than “no supply chain”: the pinned secret-scan actions, GitHub-managed Pages hosting/build, and any installer the adopter chooses are separate surfaces. Static diagrams contain no scripts or external resource references. The skill’s main risk surface remains the instructional content itself; it does not sandbox the receiving agent.

Scope

Security concerns that are in scope for this project:

• Prompt-injection risks in the methodology docs or templates — wording that could lead a downstream AI agent to treat untrusted content as instructions.
• Unsafe agent instructions — guidance that could cause an agent to take a destructive, secret-exposing, or gate-bypassing action.
• Secret leakage — credentials, tokens, or keys accidentally committed to this repository.
• Misleading or malicious workflow instructions that could harm a project adopting the methodology.

How the methodology itself defends against these is documented in methodology/13_ai_safety_and_prompt_injection.md.

Reporting a vulnerability

Please report privately rather than opening a public issue. Email polgarmiklos@gmail.com.

Include what you can:

• A description of the issue and which files or prompts are affected.
• How to reproduce it (e.g. the input that triggers unsafe behavior).
• The potential impact.
• A suggested fix, if you have one.

This is a lean, solo-maintained reference artifact (see STATUS.md): reports are read, but there is no response SLA. Critical issues — anything that could cause downstream harm — will be prioritized.

How to verify the repository yourself

You do not have to take the statements above on trust:

1. Inspect the tree. Confirm there is no package.json, no source files, no install scripts — only markdown and the one read-only workflow.
2. Run a secret scan. gitleaks detect --no-banner should report zero findings. (The repo runs this on every push and pull request.)
3. Review the safety doc. methodology/13_ai_safety_and_prompt_injection.md states the rules the methodology gives downstream agents; the templates (templates/CLAUDE.md, templates/AGENTS.md) carry the same rules in the form agents actually load.

A note for adopters

This methodology governs which instructions an AI agent obeys. It does not sandbox your agent’s runtime or scan your project’s dependencies — those remain your project’s responsibility, layered on top. Adopting the methodology is a strong default for safe AI-assisted development, not a substitute for your own security controls.

This site is open source. Improve this page.