With this solution to explicitly approve which credentials it can access for each task and the passwords themselves are never shared...Claude just can ask
1Password to fill them in for it at the right time in the browser. This is much safer.
We did some novel research earlier this year if you really want to see why it's bad for AI agents to just be able to grab things in your password manager...
https://1password.github.io/SCAM/replays/phish-shared-doc.ht... that's the nightmare scenario.