Skip to content

Security: operatoruplift/cold

Security

SECURITY.md

Security

Reporting a vulnerability

Please report it privately through GitHub's "Report a vulnerability" on this repository (Security, Advisories), not in a public issue.

Keeping your setup safe

  • Secrets live in .env.local only on your computer. Mailbox URLs carry passwords and the Twenty key opens your CRM. .env.local, .env and cold.yaml are gitignored; keep them that way. On Vercel, what you save in Settings is stored in the database encrypted (AES-256-GCM) with COLD_SECRET_KEY, which lives only in Vercel's environment settings; a copy of the database alone does not reveal it.
  • Use app passwords, not account passwords, for Gmail, Google Workspace and Microsoft 365 mailboxes, and revoke them when a mailbox is retired.
  • Scope the Twenty key. Give it a role that can read people, companies and the dialer's objects and write coldTouches and agencyScripts. It does not need workspace settings.
  • Reacher is for your machine or private network. If you expose it, set RCH__HEADER_SECRET on the server and REACHER_SECRET here.
  • The app has one login: yours. Everything under /app/ and every API call except signing in and the waitlist needs the owner's session. Use a long password (12 characters is the minimum; a short sentence is better). Settings lists every signed-in device and can sign the others out; bun run cold owner resets the password from the server.
  • On your computer bun run ui listens on 127.0.0.1 only. Anywhere, the server refuses requests whose Host or Origin is not its own and requires its own header on every API call, so other websites open in your browser cannot drive it.
  • Hosted, use HTTPS. Set COLD_PUBLIC_URL to an https:// address so the session cookie is Secure and browsers get HSTS (on Vercel this is automatic). Keep the /data volume private: it holds .env.local (mailbox passwords, API keys) and the database. On Vercel, the cron route answers only to CRON_SECRET, and sign-in limits are counted in the database, so every copy of the function shares them.
  • The Anthropic key is billed to you. Keep ANTHROPIC_API_KEY in .env.local or the server's secrets, set a spend limit in the Claude Console, and rotate it if it leaks. Settings shows what the AI features used in the last 30 days.
  • AI never sends. Research and drafts are suggestions: an email only goes out through a campaign send you start, with your caps, window and opt-outs; DMs and calls are always yours.
  • Diligence reads only the public internet. A prospect's website comes from a list anyone could have written, so before COLD reads it, every address the name resolves to must be public (private, loopback, link-local, carrier and reserved ranges are refused, IPv4 and IPv6, in every spelling), each redirect is checked again, only http and https on the normal ports are read, and every page has a time and size limit. It honours robots.txt. A name that resolves to a public address when checked and a private one when fetched (DNS rebinding) is not fully ruled out; on a server with sensitive internal services, run diligence where they aren't reachable.
  • Decision models get no contact details. Jev or Claude see the list row without email addresses, phone numbers or record ids. The TypeSafe and SubQ keys are billed to you, like the Anthropic key; Settings and the Score tab show usage per provider.
  • Prospect lists are personal data. prospects*.csv is gitignored; the only CSV in the repository is the fictional example. The waitlist keeps a salted hash of each signup's address, never the address itself.

There aren't any published security advisories