Nice article, although the title is slightly off. The private key is not compromised (as described agent forwarding uses a challenge-response method) and the attacker can only initiate connections as long as the agent is being forwarded.