🛡️ VibeGuard

Your AI coding agent writes fast. It also commits AWS keys at 2am.

VibeGuard is a pre-commit hook and GitHub Action that stops AI-generated code from shipping secrets, hardcoded API keys, and security holes — before they ever reach your repo.

---

The problem

AI coding agents are incredible — and they have zero survival instinct. They'll happily:

• Hardcode your Stripe live key "just to get it working"
• Commit a .env with production credentials
• Write SQL with string concatenation and call it done

VibeGuard sits between your agent and your repo and says no.

Quickstart

pip install vibeguard-secrets
vibeguard init        # writes .vibeguard.toml

Add to your pre-commit config:

repos:
  - repo: https://github.com/Shifu34/vibeguard
    rev: v0.1.0
    hooks:
      - id: vibeguard

pre-commit install

That's it. The next time your agent stages a secret, the commit is blocked:

VibeGuard found 2 potential secret(s):

  HIGH   src/payments.py:14  stripe-live-key
         Stripe live secret key
         stripe.api_key = "sk_live_4eC39HqLyjW..."

  MEDIUM src/config.py:3  high-entropy-secret
         High-entropy value assigned to 'api_token' (possible hardcoded secret)

Remove the secret, or allowlist the path in .vibeguard.toml

What it catches

Rule	Severity	Example
AWS access / secret / session keys	high/high/medium	AKIAIOSFODNN7EXAMPLE
Azure storage keys, GCP service-account keys	high	AccountKey=..., "type": "service_account"
Stripe live, restricted & test keys	high/high/medium	sk_live_..., sk_test_...
Shopify admin API tokens	high	shpat_...
Twilio, SendGrid, Mailgun keys	high	SK..., SG..., key-...
GitHub tokens (classic, fine-grained, OAuth, app, server)	high	ghp_..., gho_..., ghs_...
GitLab PATs, npm / PyPI tokens, npmrc auth tokens	high	glpat-..., npm_..., _authToken=...
Heroku, DigitalOcean, Cloudflare tokens	high/high/medium	dop_v1_...
Slack tokens & webhooks	high	xoxb-...
Discord bot tokens & webhooks	high	discord.com/api/webhooks/...
Telegram bot tokens	high	<bot-id>:<35-char token>
Sentry auth tokens, Linear API keys	high	sntrys_..., lin_api_...
OpenAI, Anthropic, Hugging Face, Google keys	high	sk-..., sk-ant-..., hf_..., AIza...
Private key blocks & PuTTY key files	high	-----BEGIN RSA PRIVATE KEY-----, PuTTY-User-Key-File-2:
DB connection strings with credentials	high	postgres://admin:s3cret@...
High-entropy assignments	medium	api_token = "a9F3kQ7z..." (no known prefix needed)

40 rules total, plus a generic secret-assignment pattern for the long tail.

Plus an optional LLM review that reads the actual diff and flags what regexes can't: SQL injection, auth bypass, insecure crypto, SSRF, path traversal. Enable it with:

[llm]
enabled = true
model = "gpt-4o-mini"   # any OpenAI-compatible endpoint works

export VIBEGUARD_API_KEY="..."

GitHub Action

Scan every PR diff automatically:

- uses: Shifu34/vibeguard@v0.1.0
  with:
    base: origin/${{ github.base_ref }}
    # llm-review: "true"
    # api-key: ${{ secrets.VIBEGUARD_API_KEY }}

Configuration

vibeguard init writes a .vibeguard.toml — all knobs in one place:

fail_on = "high"   # "high" or "medium"

allowlist = [
  "*.md",
  "docs/**",
  "tests/**",
]

[llm]
enabled = false
model = "gpt-4o-mini"
# base_url = "https://api.openai.com/v1"

Manual scans:

vibeguard scan            # staged changes (what's about to commit)
vibeguard scan --all      # every tracked file
vibeguard scan --base origin/main   # diff against a branch (CI)
vibeguard scan --format json        # machine-readable output
vibeguard scan --format sarif       # SARIF 2.1.0 (e.g. for GitHub code scanning)

Baselines

Already triaged some findings and don't want CI to keep failing on them? Record a baseline — later scans only report new secrets:

vibeguard scan --all --update-baseline   # record current findings
vibeguard scan --all --baseline .vibeguard-baseline.json  # only new findings fail

Or set it once in .vibeguard.toml and forget it:

baseline = ".vibeguard-baseline.json"

Fingerprints are based on the rule, file, and matched content (not line numbers), so a known secret that moves within a file stays suppressed. Commit the baseline file so the whole team and CI share it.

Why not gitleaks / trufflehog?

Those are excellent secret scanners — VibeGuard happily stands on their shoulders conceptually. The difference:

• Built for the agent era: the threat isn't a tired dev, it's a tireless agent committing at machine speed. VibeGuard is optimized for pre-commit speed and agent workflows.
• LLM diff review: catches logic-level vulnerabilities (injection, auth bypass), not just known secret formats.
• 60-second setup: one pip install, one pre-commit block, zero config required.
• Zero dependencies: the core scanner is stdlib-only Python.

Roadmap

• More secret rules (Twilio auth tokens, Vault tokens, ...)
• .env and config-file aware scanning
• VS Code / JetBrains extensions
• vibeguard --fix: auto-move secrets to env vars

Contributing

PRs welcome — especially new secret rules with tests. See tests/test_secrets.py for the pattern: one rule, one test, one line of regex.

pip install -e ".[dev]" 2>/dev/null || pip install -e .
python -m pytest

License

MIT — go build something safe with it.

---

⭐ If VibeGuard saves you from one leaked key, give it a star. That's the whole business model.