Despite extensive research, covert channels are a principal threat to information security. Covert channels employ specially-crafted content or timing characteristics to transmit internal information to external attackers. Most techniques for detecting covert channels model legitimate network traffic. However, such an approach may not be applicable in dynamic virtualized environments because traffic for modeling normal activities may not be available.This paper describes Observer, a real-time covert channel detection system. The system runs a secure virtual machine that mimics the vulnerable virtual machine so that any differences between two virtual machines can be identified in real time. Unlike other detection systems, Observer does not require historic data to construct a model. Experimental tests demonstrate that Observer can detect covert channels with a high success rate and low latency and overhead.