AI commits secrets every week. Here's how to stop it.

GitHub's secret scanner finds tens of thousands of leaked API keys per day. A huge fraction come from AI tools committing .env files, hardcoding api_key = "sk-..." into source, or pushing scratch notebooks with real credentials inside. If you ship anything this year, you'll write to a private API. The single most important habit in this whole course is making sure your keys never end up in git.

The pattern that works

Three rules. They compound:

1. Keys live in environment variables, never in source code. Your Python code reads them via os.getenv("ANTHROPIC_API_KEY"). The actual value lives in a .env file you load at startup, or in the platform's secret manager (Vercel, Railway, GitHub Actions).
2. .env is gitignored on day one. Before you write any code, the very first line in your .gitignore should be .env. Then .env.local, .env.*.local, and friends.
3. There's a .env.example checked in. A template with the same keys but empty or placeholder values, so a teammate (or future you) knows what variables the app needs.

The minimal .gitignore for any AI project

# secrets
.env
.env.*
!.env.example

# python
__pycache__/
*.pyc
.venv/

# editor
.DS_Store
.vscode/
.idea/

The !.env.example line says "ignore .env-everything except the example template." The exclamation mark is an un-ignore.

The pattern AI ships every time

import os
from anthropic import Anthropic

# python-dotenv reads .env into os.environ at startup
from dotenv import load_dotenv
load_dotenv()

api_key = os.getenv("ANTHROPIC_API_KEY")
if not api_key:
    raise RuntimeError("ANTHROPIC_API_KEY not set — copy .env.example to .env")

client = Anthropic(api_key=api_key)

Five lines. Same pattern for OpenAI, Stripe, GitHub tokens, database connection strings — anything sensitive.

What to do when a key leaks (and it will)

You will leak a key. Everyone has. The fix is fast if you act fast:

1. Rotate the key immediately. Open the provider's dashboard, revoke the old key, generate a new one. The leaked key is now useless to anyone who scraped it. Do this first, before cleaning up the repo.
2. Update your local .env with the new key. Test that the app still works.
3. Remove the key from git history. git rm the file and commit. If the key was pushed, the secret is in history forever on GitHub — rotation in step 1 is what actually protects you. Tools like git filter-repo can scrub history, but the key already being public is the threat, not the file existing.
4. Audit usage. Check the provider's logs for any calls made between the leak and the rotation.

The rotation in step 1 is the only thing that actually matters. The other steps are hygiene.

Where AI specifically gets this wrong

• Hardcodes the key in a Jupyter cell. "Just for this notebook." Then commits the notebook. The key is now public.
• Prints the full key to stdout for debugging. Logs end up in cloud log services that get indexed.
• Adds .env to git "by accident" with git add .. Always git status first.

Browser note: there's no real os.environ in here, and python-dotenv isn't installed. We'll use a Python dict to simulate environment variables. The pattern (getenv, default values, missing-key handling) is identical.

Run the editor. We load a (fake) key from a (fake) env and print only a redacted preview.