2.5 Private Addressing and NAT
2.5.2 Discontiguous subnets

Mixing private addresses with globally unique addresses can create discontiguous subnets, which are subnets from the same major network that are separated by a completely different major network or subnet.

In the figure, Site A and Site B both have LANs that are addressed using subnets from the same major net (207.21.24.0). They are discontiguous because the 10.0.0.4/30 network separates them. Classful routing protocols, notably RIPv1 and IGRP, cannot support discontiguous subnets because the subnet mask is not included in routing updates. If Site A and Site B are running RIPv1, Site A will receive updates about network 207.21.24.0/24 and not about 207.21.24.32/27 because the subnet mask is not included in the update. Because Site A has an interface directly connected to that network (in this case, E0), Site A will reject Site B's route.

Even some classless routing protocols require additional configuration to solve the problem of discontiguous subnets. RIPv2 and EIGRP automatically summarize on classful boundaries unless explicitly told not to. Usually, this type of summarization is desirable, but in the case of discontiguous subnets, the following command must be entered for both RIPv2 and EIGRP to disable automatic summarization: 

Router(config-router)#no auto-summary

Finally, when using private addresses on a network that is connected to the Internet, you should filter packets and routing updates to avoid "leaking" any RFC 1918 addresses between autonomous systems. For example, if both you and your provider use addresses from the 192.168.0.0 /16 block, your routers could get confused if confronted with updates from both systems.