I definitely think we'll write tools to analyse the permissions and explain the worst case outcomes.

I can accept burning tokens and redo on the scale of hours. If I'm losing days of effort I'd be very dissatisfied. Practically speaking people accept data loss because of poor backups, because backups are hard (not technically so much as administratively), but I'd say backups are about to become more important. Blast limiting controls will become essential -- being able to delete every cloud hosted photo is just a click away. Spinning up thousands of EC2 nodes is incredibly easy, and credit cards have extremely weak scoping.

The LLM will easily leak these credentials out. So the creds should be outside the sandbox, and the only thing the sandbox should see is a connection API that opens a socket/file handle.

Alternatively where is needs an API key, it should be one bound to the endpoint using it. E.g. a ticket granting ticket is used to create a bound ticket.

A copy on write filesystem would be an interesting way to sandbox writes, but there is difficulty in checking the diff.