The ssh transaction/repl log thing is frightening to me. I get it, but when I've needed to get creds for a user to log into a service, one of my go-tos has been to pull shell history on all the hosts they log into since they've almost definitely typed their password prematurely at some point. So, logging all that sounds like a really fun way of making a password store!
The concern being raised here is that transcript-level SSH audit logs are the equivalent of permanent shell histories for everyone, and they are. But if you're giving team members a reason to ever type a password into an SSH session, you're got a bigger gap to close. We already have to do secrets management at scale, because it's a feature we provide to our customers, and so we already have a process for loading secrets into environments for host work.
I'd be more worried (but not terrified) that the session transcripts can teach an attacker a fair bit about how systems work, should an attacker get access to those. Of course only a small subset of attackers is going to care...
