Is this not a good use case for AI in your email client (local-only to avoid more opportunities for data to leak)?

Have the client-embedded AI view the email to determine if it contains a link to a purported service. Remotely verify if the service URL domain is valid, by comparing to the domains known for that service

If unknown, show the user a suspected phishing message.

This will occasionally give a false positive when a service changes their sending domain, but the remote domain<->service database can then be updated via an API call as a new `(domain, service)` pair for investigation and possible inclusion.

I feel like this would mitigate much of the risk of phishing emails slipping past defenses, and mainly just needs 2 or 3 API calls to service once the LLM has extracted the service name from the email.

No, the solution to a security problem is not to radically increase the vulnerable attack surface.