Evaluating Control Protocols for Untrusted AI Agents