It's a real threat, albeit an exotic one. The threat model is a malware compromised host, with a wallet (possibly a low power hardware wallet like a Trezor) that can understand the payment protocol and sign transactions, but maybe not do a whole lot more than that. For instance, probably it cannot do HTTPS connections itself. So a virus on the host could swap the refund address for one that is owned by the attacker, and then try to make the merchant issue an automatic refund, thus bouncing the funds back off the merchant to the them.

If there are merchants that offer large, automatic refunds, it could be an issue. I'm not sure how common that might be in reality. Steven or Tony would know. Timo's protocol is an interesting solution, but again, at this point the feature set for v1 is pretty much locked down.