Your AI Coder Writes Fast. It Also Writes API Keys Into Source Files.

An open-source skill for Cursor, Claude Code, Copilot, and any agent that reads AGENTS.md.

The leak that happens before git push

If you use AI coding agents, you’ve probably seen this:

stripe.api_key = "sk_live_..."

Or a Postgres URL with a real password. Or an OpenAI key in a .env the agent created and almost committed.

The agent isn’t malicious. It’s pattern-matching from training data and “finishing” config files the way tutorials do — with inline secrets.

Traditional tools catch this later:

Tool: When it runs

GitHub secret scanning: After push

gitleaks / truffleHog: After commit (or in CI)

Pre-commit hooks: At commit time

All useful. None of them run inside the agent while the model is still generating the file.

That gap is what leash-secrets fills.

---

What leash-secrets does

leash-secrets is an AI agent skill — a protocol your coding agent follows on every edit:

1. SCAN every line for known secret patterns

2. CLASSIFY as critical, warning, or safe (placeholders like changeme pass through)

3. BLOCK critical findings — show redacted value, explain risk, provide fix steps

4. ACT — replace hardcoded secrets with env vars, update .env.example, check .gitignore

When it fires, the output looks like this:

⛔ LEASH-SECRETS — SECRET DETECTED
━━━━━━━━━━━━━━━━━━━━━━━━━
Type:  Stripe Live Secret Key
File:  payments.py:3
Value: sk_liv....9iU
Risk:  Can create charges and access customer payment data.
━━━━━━━━━━━━━━━━━━━━━━━━━
FIX:   Use STRIPE_SECRET_KEY env var + .env.example entry

The agent stops. The key doesn’t make it into the repo.

---

Where it runs

Install once; it wires into whichever agents you use:

• Cursor — .cursor/rules/leash-secrets.mdc

• Claude Code / Codex / Cline — skills/ or AGENTS.md

• GitHub Copilot — .github/copilot-instructions.md

• Gemini CLI, Windsurf — supported via the universal installer

Install (macOS / Linux / WSL):

curl -fsSL https://raw.githubusercontent.com/FasterApiWeb/leash-secrets/main/scripts/install.sh | bash

CLI only:

npm install -g leash-secrets
leash-secrets scan .

Patrol mode is on by default — the agent checks as it writes. A pre-commit hook is there as a second line of defense.

---

Coverage

71 patterns across 11 provider categories:

• Cloud: AWS, GCP, Azure

• AI: OpenAI, Anthropic, Hugging Face, Cohere, Replicate

• Payments: Stripe, PayPal, Square

• Databases: PostgreSQL, MySQL, MongoDB, Redis, Supabase, PlanetScale

• Messaging: Slack, Discord, Twilio, SendGrid

• CI/CD: npm, PyPI, Docker Hub, Vercel, Terraform

• Crypto: RSA, OpenSSH, EC, PGP private keys

• Generic: passwords, JWT secrets, Bearer tokens

Patterns are JSON — add your own (patterns/my-company.json) and contribute upstream.

---

Benchmarks (reproducible, not hand-wavy)

I’m skeptical of security tools that claim “94% accuracy” with no corpus. So leash-secrets ships a published benchmark:

• 41 labeled cases in benchmarks/corpus/cases.json

• Run locally: npm run benchmark:corpus

Last results (pattern scanner):

Metric Result Recall (30 positives) 100% False positives (8 negatives) 0% Known documented gaps 3

Those 3 gaps are intentional honesty: multiline secrets, base64-only blobs, truncated tokens. We document them instead of hiding them.

This measures the regex scanner — reproducible and verifiable. I’m not publishing unverified “LLM prompt compliance” numbers.

---

How it fits your stack

leash-secrets does not replace gitleaks, GitHub secret scanning, or your SAST pipeline.

Think of it as defense layer 0:

Agent writes code → leash-secrets blocks secrets
       ↓
File saved → pre-commit hook (backup)
       ↓
Commit → gitleaks / CI
       ↓
Push → GitHub secret scanning

Each layer catches what the previous one missed.

---

Project status

• Open source (MIT)

• npm: leash-secrets@1.2.1

• Docs: https://fasterapiweb.github.io/leash-secrets

• Repo: https://github.com/FasterApiWeb/leash-secrets

• Maintained by EshwarCVS / FasterApiWeb

Good first issues are open if you want to add patterns (Datadog, Sentry, Cloudflare) or stack-specific guides.

---

Try it

1. Run the installer (takes ~30 seconds)

2. Ask your agent to “add Stripe payments” or “connect to Postgres”

3. Watch whether it reaches for env vars instead of hardcoding keys

If you hit false positives or want a provider we don’t cover yet, open an issue. This is early — feedback shapes what ships next.

---

Links

• GitHub: https://github.com/FasterApiWeb/leash-secrets

• npm: https://www.npmjs.com/package/leash-secrets

• Benchmark corpus: https://github.com/FasterApiWeb/leash-secrets/tree/main/benchmarks

If you use leash-secrets in a project or paper, cite via CITATION.cff.