Security checks

Keep .env files out of version control. Commit an example file with placeholders instead.

Run a secret scanner before committing and in CI. Include Git history in the scan.

Revoke and replace exposed credentials. Removing them from a file or Git history does not invalidate them.

Check that browser JavaScript, public assets, and API responses contain no private keys.

Use explicit dependency versions and commit the lockfile for reproducible installs.

Verify authentication on the server for every endpoint that exposes private data or actions.

Try accessing another user’s records by changing IDs in URLs and requests. Confirm that access is denied.

Where row-level security is used, verify read and write policies with different users, roles, and tenants.

Use a maintained authentication library or provider instead of designing your own login and password handling.

Verify that sessions can be invalidated. Use appropriate token lifetimes and test logout and account revocation.

Enforce privileged actions on the backend. Hiding an admin button is not an authorization check.

Rate-limit login, signup, and password-reset endpoints to reduce brute force and abuse.

Validate types, lengths, allowed values, and request sizes on the server. Client validation can be bypassed.

Use bound query parameters. Never concatenate untrusted input into SQL queries.

Escape output for its HTML, URL, or JavaScript context. Sanitize any user HTML that you intentionally allow.

Allow only the origins that need browser access. Keep authentication and authorization checks in place independently.

Make buckets and uploaded objects private by default. Grant access only to the users who need it.

Validate file types and sizes, store uploads away from executable application code, and isolate file processing.

Verify the provider’s signature before processing a webhook. Reject invalid requests and handle replayed events safely.

Enforce usage budgets and a cutoff for paid AI calls. Alerts alone do not stop spending.

Apply per-user or per-tenant request and concurrency limits to prevent credit draining.

Account for prompt injection in user input, retrieved documents, and tool results. Keep permission checks outside the model.

Limit tool access and permissions, especially SQL and shell execution. Require approval for sensitive actions.

Check package names, sources, and maintainers before installing dependencies suggested by a model.

Review agent instructions, skills, and MCP connections. Check what they can execute and which data they can access.

Keep production credentials out of agent prompts and workspaces unless explicitly needed. Grant only the required access.

Return generic public errors. Keep detailed diagnostic information in restricted internal systems.

Remove credentials, tokens, and unnecessary personal data from application logs and traces.

Record who performed sensitive actions, what changed, and when. Restrict access to audit records.

Restore a backup in an isolated environment and verify the recovered data. A successful backup job is not a restore test.