Bot-like malware has posed an immense threat to computer security. Bot detection is still a challenging task since bot developers are continuously adopting advanced techniques to make bots more stealthy. A typical bot exhibits three invariant features along its onset: (1) the startup of a bot is automatic without requiring any user actions; (2) a bot must establish a command and control channel with its botmaster; and (3) a bot will perform local or remote attacks sooner or later. These invariants indicate three indispensable phases (startup, preparation, and attack) for a bot attack. In this paper, we propose BotTracer to detect these three phases with the assistance of virtual machine techniques. To validate BotTracer, we implement a prototype of BotTracer based on VMware and Windows XP Professional. The results show that BotTracer has successfully detected all the bots in the experiments without any false negatives.