Cursor: rules, ignore files and safe agent settings

Last reviewed October 2026 · Sources: Cursor docs: Rules, Ignore files, Agent terminal, Cursor security

Coding agents move fast and do what you ask. These settings make them safer by default.

1. Add a security rule the agent always follows

Cursor supports project rules (in .cursor/rules/) that are added to the agent's context. Add one for security:

---
description: Security defaults for this project
alwaysApply: true
---
- Never put API keys, tokens or passwords in source code or client-side files. Use environment variables loaded on the server.
- Every API route and server action must check the user is authenticated AND authorized for the specific record.
- Validate all user input on the server (types, length, allowed values).
- Use parameterized queries or the ORM; never build SQL with string concatenation.
- Don't disable security checks, RLS, or CORS restrictions to "make it work". Ask me first.
- When adding a dependency, prefer well-known packages and tell me the exact name.

2. Keep secrets out of the agent's reach

• Add .env*, key files and credential folders to .cursorignore (see Cursor's ignore-file docs for how this limits what's indexed and read).
• Keep the same files in .gitignore so they never reach GitHub.
• If a key ever appeared in a commit, rotate it. Deleting the commit isn't enough.

3. Decide what the agent may run without asking

Cursor's agent can run terminal commands. Review the auto-run settings in Cursor's agent/terminal docs: keep approval on for commands that delete files, deploy, push, or touch production databases, and use allow/deny lists where available.

4. Check privacy settings for your situation

Review Cursor's privacy mode and data-handling settings (see Cursor security) if you work with client code or sensitive data.

5. Review before you ship

Act as a security reviewer. List every route, server action and database query in this
repo. For each: who can call it, what it reads/writes, and whether input is validated
and the caller is authorized. Flag anything that exposes secrets to the browser.

Then fix and re-check. The agent can miss things, so verify the critical paths yourself.