Safe Use of AI Coding Assistants (Secrets, Licences, Data)

In short: never paste secrets or .env files, check the licence of suggested code, know what leaves your machine (cloud vs local), follow company policy, and never auto-approve dangerous shell commands.

Why safe use matters

One pasted API key in a cloud chat can bill your cloud account overnight. One copy-pasted snippet under a strict open-source licence can create a legal problem for your company. One agent command like rm -rf can wipe a folder. Safety is part of the skill, not an extra.

What the risks are

• Secrets: API keys, passwords, .env files, private keys, customer data, production connection strings.
• Licences: code the AI suggests may match open-source projects. Some licences force you to open your whole product (copyleft). Others are fine for commercial use (MIT, Apache-2.0).
• Data leaving your PC: cloud assistants send prompts and often parts of your files to a company server. Local tools like Ollama keep the model on your machine.
• Dangerous actions: agent mode can run shell commands, edit many files, install packages, or push to git if you allow it.
• Company policy: many Indian IT companies ban pasting client code into public AI tools, or allow only approved tools.

How a safe session flows

Rule 1 — do not paste secrets

1. Never put .env, credentials.json, PEM keys, or database passwords in a prompt.
2. In Claude Code, deny those paths in .claude/settings.json (example below).
3. In Copilot / Cursor / Continue, keep secrets out of the workspace the tool indexes, or add them to ignore lists.
4. If you pasted a key by mistake: rotate it now (create a new key, delete the old one) in AWS, GitHub, Stripe, etc.

{
  "permissions": {
    "deny": [
      "Read(./.env)",
      "Read(./.env.*)",
      "Read(./secrets/**)",
      "Read(**/*.pem)",
      "Bash(rm -rf:*)",
      "Bash(git push --force:*)"
    ]
  }
}

Rule 2 — check licences of suggested code

1. If the AI pastes a large block that looks copied from a library, ask: “Is this from an open-source project? Which licence?”
2. Search a unique line on GitHub. If it matches a GPL or AGPL file and your product is closed-source, do not use it as-is.
3. Prefer asking the AI to write a small function from a clear spec, not “copy how library X does it”.
4. When you use a real library, add it through your package manager (pip, npm, Maven) so the licence is tracked.

Rule 3 — know what leaves your machine

Rule 4 — company policy first

1. Ask your lead: which AI tools are allowed? Is client code allowed in any cloud tool?
2. Use only the approved account (often a company Copilot or Claude seat), not your personal free account for office work.
3. Do not paste customer names, phone numbers, Aadhaar, PAN, bank data, or medical records into any AI.
4. If policy is unclear, use a local model or write the code yourself until someone confirms.

Rule 5 — never auto-approve dangerous shell

1. Keep permission prompts on in Claude Code, Copilot Agent, Cursor Agent, Devin Local, Kiro.
2. Read every command before Yes. Refuse: rm -rf, mkfs, dd, force-push to main, curl-piped-to-bash from unknown URLs, disabling firewalls.
3. Prefer Plan mode first for big tasks: the tool lists steps; you approve; then it runs.
4. On a production server or a shared machine, do not run agent mode at all unless your team has a written process.

A one-page safe checklist

Quick quiz

Save tokens and money

Safety also saves money: smaller prompts without secrets and logs mean fewer tokens. Put standing deny rules in one short settings file instead of repeating “do not read .env” in every chat.

Save these cards