Draconian firewall rules. If you know how the application is vulnerable, then you can write rules to expect those responses from the application and nothing else and to limit how far the known vulnerabilities go, while still gathering the traffic data. You could even write rules so that expected 'bad' responses from the server are redirected to a black hole or another server so that you can capture that traffic, too.
I would also make sure that the virtual machines and server do not have access to other VM/servers or networks.
As Legolas says, you are setting up a high interaction honeypot, so set it up like that to begin with, but you need to limit the actual malicious responses that your server creates.