xChe AI-App Security Pack

License: MIT PRs Welcome Made for Claude Code Status: Active

Free, opinionated security guardrails for apps built with AI coding tools — Claude Code, Cursor, Lovable, Bolt, Codex — so the speed doesn't ship you a breach.

Drop-in CLAUDE.md rules, a full OWASP / API / LLM / MCP security policy, plus secret-scanning hooks and a CI gate. Hardens your AI coding setup against the exact bug classes behind every "vibe-coded app leaked its database" headline — before the code is written.

Why this exists

AI coding tools generate working code, not secure code. The result is now a recurring headline:

  • A vibe-coded AI social network (Moltbook) left its database wide open — ~1.5M API tokens and 35K emails exposed — because the generated code skipped Row-Level Security. (Wiz)
  • A viral AI-built recovery app (Quittr) leaked the sensitive personal data of hundreds of thousands of users — including ~100K minors — via a misconfigured Firebase database. (404 Media)
  • Independent studies keep finding a large share of AI-generated code ships insecure — ~40% in NYU's GitHub Copilot study, ~45% across languages in Veracode's 2025 GenAI Code Security report. (NYU)

The bugs are almost always the same handful: hardcoded secrets, broken access control, missing input validation, exposed databases, no security headers. This pack hardens your setup against exactly those.

What's inside

File What it does
AGENTS.md Condensed rules in the cross-tool standard — read by Cursor, Codex, Copilot, Gemini CLI, Aider, Windsurf & 25+ agents.
CLAUDE.md The same rules + Claude-specifics — drop into your global ~/.claude/CLAUDE.md so every project inherits them.
SECURITY.md The full policy — OWASP Top 10 + API Top 10 + LLM/MCP Top 10 + cloud/infra rules.
plugins/ai-app-security/ A Claude Code plugin: a secure-coding skill (applies the rules while building) + an /audit skill (manual full review) + a hook that blocks secret commits. One-command install.
.pre-commit-config.yaml A git pre-commit hook that scans for secrets (gitleaks).
.github/workflows/security.yml A CI gate: secret scan + dependency audit on every push/PR.

Install as a Claude Code plugin (one command)

/plugin marketplace add xChechi/xche-ai-app-security-pack
/plugin install ai-app-security@xche

This activates three things:

  • a secure-coding skill that auto-applies the security rules whenever Claude Code writes or edits auth / database / API / input / secrets / LLM-MCP code (no command needed — it triggers on the task);
  • an /ai-app-security:audit command for a full manual review of selected code or recent changes;
  • a best-effort secret guard that runs when Claude Code itself makes a git commit (uses gitleaks if installed, skips otherwise). It's a convenience — not a complete gate; see the note below.

For the strongest, always-on coverage, also drop CLAUDE.md into ~/.claude/CLAUDE.md (or AGENTS.md into your repo). A skill triggers when the task matches; CLAUDE.md is in context on every turn. Use both.

How the secret nets actually layer (honest version)

The plugin's secret guard is a best-effort check that runs when Claude Code itself makes a git commit. It is not a complete gate: it doesn't cover commits from your own terminal or IDE, and it can miss git commit -am (those changes aren't staged yet at hook time). Treat it as early convenience, not a guarantee.

The three nets are not equal — install all of them, but know what each does:

Net Covers Role
pre-commit hook (.pre-commit-config.yaml) all local commits, any tool; catches -am (runs at the real commit moment) Primary local gate
CI workflow (.github/workflows/security.yml) every push/PR; can't be bypassed or skipped Hard backstop
Claude Code plugin hook only commits Claude Code itself runs Convenience — catches the common "Claude committed a key" case early

If you only do one thing, install the pre-commit hook. The CI gate is what ultimately stops a secret reaching the remote.

Or use the files directly (any tool, 2 minutes)

  1. Cross-tool rules: copy AGENTS.md into your repo root (Cursor/Codex/Gemini/etc. read it automatically).
  2. Claude Code global rules: copy CLAUDE.md's contents into ~/.claude/CLAUDE.md.
  3. Per-project policy: drop SECURITY.md into your repo root for the full reference.
  4. Block secret leaks locally: install pre-commit, then pre-commit install.
  5. Block them in CI: copy .github/workflows/security.yml into your repo.

Works with

AGENTS.md covers the cross-tool agents; CLAUDE.md + the plugin cover Claude Code natively; and the policy, gitleaks hooks, and CI gate are tool-agnostic — use them with Cursor, Codex, Gemini CLI, Bolt, Lovable, or any AI-assisted (or human) codebase. AGENTS.md and CLAUDE.md are kept in sync and are each self-contained, so either works dropped in on its own.

⚠️ This is a strong baseline, not a guarantee. It catches the common classes; business-logic flaws and novel issues still need human review. Pair it with /security-review and — for anything handling money, auth, or personal data — a real audit.

Shipped something fast and want it checked?

If you've built an app with an AI tool and want to know it won't be the next breach headline, I do security audits + remediation for AI-built apps.

→ stefannasev.dev · LinkedIn

Contributing

Found a gap or a rule worth adding? PRs and issues welcome — this policy improves with real-world findings.

⭐ If this helped, star the repo — it helps other AI builders find it before they ship a leak.

License

MIT — use it, fork it, ship it.