Cursor applies best-effort redaction for known secret patterns before a transcript is shared. What does that leave the person sharing it responsible for?

The principle — Redaction before sharing is best-effort. It matches known secret patterns, and it can miss secrets, especially custom values or values without a recognisable prefix.

Why the key is correct — Pattern matching finds what it has seen before. An internal token in a bespoke format, a password typed into a prompt, or a connection string built at runtime gives the matcher nothing to lock onto, so it survives into the published copy. Only your own read catches it.

Why the others are wrong — Choosing the audience scrubs nothing from the content. Redaction is applied to what gets published, not only to your local view, so wholesale rotation is not the remedy. And nothing waits for you to flag messages by hand.

Remember this — A public link cannot be unpublished from anyone who has already copied it.

Sources — Cursor's Shared transcripts documentation.