If the history audit turns up a real credential, rotate it immediately. A key that appeared in any git commit should be treated as compromised even if the repository is private.

Harden Your Prompts, Not Just Your Code

You can reduce the frequency of AI-generated credential leakage by changing how you write prompts. Be explicit:

• Avoid: "Write a function that calls the OpenAI API"
• Prefer: "Write a function that calls the OpenAI API. Load the API key from the environment using os.environ. Raise a clear error if the key is missing. Do not hardcode any credentials."

This works because language models follow instructions. The model will still default to the training-data pattern if you do not specify otherwise, but an explicit instruction overrides it most of the time.

Gate Secrets at the Repository Level

Prompt hygiene and manual audits are not enough on their own. Add tooling that prevents the commit from happening:

# Install pre-commit and the detect-secrets hook
pip install pre-commit detect-secrets

# .pre-commit-config.yaml
repos:
  - repo: https://github.com/Yelp/detect-secrets
    rev: v1.4.0
    hooks:
      - id: detect-secrets
        args: ['--baseline', '.secrets.baseline']

This hook scans every staged file before the commit lands. It catches the pattern before it enters history.

Key Takeaways

• AI coding assistants reproduce hardcoded credential patterns from training data. Every AI-generated file that touches external APIs should be reviewed for embedded secrets before commit.
• A git history audit catches credentials that were added and later "deleted" from the working tree. Rotation is the only safe response to a key that appeared in any commit.
• Combining explicit prompt instructions, pre-commit secret scanning, and environment-only credential loading eliminates most AI-introduced token leakage before it can be exploited.

If the history audit turns up a real credential, rotate it immediately. A key that appeared in any git commit should be treated as compromised even if the repository is private.

Harden Your Prompts, Not Just Your Code

You can reduce the frequency of AI-generated credential leakage by changing how you write prompts. Be explicit:

\n
• Avoid: \"Write a function that calls the OpenAI API\"
\n
• Prefer: \"Write a function that calls the OpenAI API. Load the API key from the environment using os.environ. Raise a clear error if the key is missing. Do not hardcode any credentials.\"
\n

This works because language models follow instructions. The model will still default to the training-data pattern if you do not specify otherwise, but an explicit instruction overrides it most of the time.

Gate Secrets at the Repository Level

Prompt hygiene and manual audits are not enough on their own. Add tooling that prevents the commit from happening:

\n# Install pre-commit and the detect-secrets hook\npip install pre-commit detect-secrets\n\n# .pre-commit-config.yaml\nrepos:\n  - repo: https://github.com/Yelp/detect-secrets\n    rev: v1.4.0\n    hooks:\n      - id: detect-secrets\n        args: ['--baseline', '.secrets.baseline']\n

This hook scans every staged file before the commit lands. It catches the pattern before it enters history.

Key Takeaways

\n
• AI coding assistants reproduce hardcoded credential patterns from training data. Every AI-generated file that touches external APIs should be reviewed for embedded secrets before commit.
\n
• A git history audit catches credentials that were added and later \"deleted\" from the working tree. Rotation is the only safe response to a key that appeared in any commit.
\n
• Combining explicit prompt instructions, pre-commit secret scanning, and environment-only credential loading eliminates most AI-introduced token leakage before it can be exploited.
\n