Yes, the systems with the log parsing bug are part of an internal subnet. As with most web scale companies HTTPS requests are terminated on a unified edge and load-balanced to web service hosts in internal clusters. In this case the malicious header was maintained in the backend requests and ended up in the application log, which triggered the command injection. Everything I wrote above is correct and is in no way incompatible with the fact that the affected machines have RFC1918 addresses.