http://serverfault.com/questions/338439/ssh-sessions-terminate-abruptly-with-mes
On-the wire level errors should be eliminated by the Ethernet and TCP checksums. It is possible that corrupted Ethernet frame / IP packet could slip now and then up to the higher levels of the networking stack, but it really isn't very probable and definitely shouldn't lead to a repeatable issue. Therefore I see two main possibilities:
1) There's somebody fooling with your transmission, modifying in-transit data so that the Ethernet/IP checksums match, but encryption layer sees rubbish, or
2) More probably, you have some hardware error, most probably CPU or RAM, that damages data after it has been taken off-the-wire. Again, ECC RAM should reduce probability of this happening, but overheating / dying CPU can play hell with data.
I don't know if TCP Offload Engine with a mis-behaving NIC/driver could produce such errors, but that line of thought could explain (changed driver?) the correlation of problems with the update.
tcp checksum offload:
http://blog.csdn.net/ixidof/article/details/6658125