The in-app assistant answers questions about your account, which means it reads some of your data and sends some of it to a model API. What it may see is now a mechanism rather than an intention.

Never sent, in either direction

• •Credentials and secrets. Anything credential-shaped in a message you type — access keys, tokens, passwords, connection strings, private keys, signed URLs — is removed before the message is stored, before it reaches the model, and before it can enter any cache or support email. Redaction happens once, at the door: scrubbing later would be theatre, because by then the value has already been written down.
• •Object contents and file names. The assistant reads account metadata — plans, limits, bucket names, job status, failure reasons — and never object keys. Where a name could leak through a written activity description or a job's error message, it is removed there too.

Minimum metadata

The assistant's tools select named columns, not rows. They never read the credentials table, and a bucket that has lost its credential is reported to the assistant as a yes/no, not as which one.

Why we are telling you this

"Here is my key, why doesn't it work" is the most natural thing in the world to paste into a support box about credentials. It should not be the thing that puts a secret in a log — so the product assumes you will, and handles it.