Skip to content

What's safe to paste into an AI tool

Most of what you’d share in a screen-share is safe to paste into a hosted AI tool. This page names the exceptions — the categories that need to stay out of the prompt, and what to do the day something slips.

Assume anything you paste into a hosted AI tool leaves your control. Paste what you’d be comfortable seeing in a support-ticket export; keep the rest out. When the only useful prompt would need something in the “keep out” list, that’s a signal to close the tab, not to redact harder.

  • Secrets and credentials. API keys, tokens, private keys, connection strings. If one lands in a chat, rotate it — deleting the message isn’t enough.
  • Customer data. Names, emails, PII, order IDs. Even one-line snippets from a production log usually carry more than the line suggests.
  • Unreleased IP. Internal-only APIs, unshipped product code, roadmap detail. Treat the paste like a public blog post.
  • Licence-restricted code. Copyleft snippets or third-party proprietary code. Both create obligations you don’t want the model quoting back into your own diff.
  • Redacting only the obvious. [email protected] looks fine; the stack trace above it still names the customer.
  • Trusting “private” chats. Read the tool’s retention and training terms before you rely on the label.
  • Assuming you’ll remember. Once a secret is in a chat, treat it as leaked and rotate — even if you deleted the message.
  1. Rotate the secret or revoke the token.
  2. Delete the message where you can; assume the model still saw it.
  3. Note the workflow that let it in. The second incident is the one that matters.