I see it doing the standard recon and working the appropriate exploit that would likely provide the most leverage with the least noise, but from within the target environment. So, yea, you've provided some good examples.

Once behind the firewall, outbound connections to a c&c could set off alarms. Especially if the connections are sourced from multiple internal systems to a single external one. So let an ai decide what to do once lateral movement options become available, then report back findings at a more appropriate time and method that is also less likely to ring bells.

That's just a quick synopsis. There is a lot more that could be done, but the secondary exploitation/invasion stages would probably see the biggest gains.