In this paper, we address issues related to defending against wide-spreading worms on the Internet. We study a new class of worms called the self-adaptive worms. These worms dynamically adapt their propagation patterns to defensive countermeasures, in order to avoid or postpone detection, and to eventually infect more computers. We show that existing worm detection schemes cannot effectively defend against these self-adaptive worms. To counteract these worms, we introduce a game-theoretic formulation to model the interaction between worm propagator and defender. We show that the effective integration of multiple defensive schemes (e.g., worm detection, forensics analysis) is critical for defending against self-adaptive worms. We propose different combinations of defensive schemes for different kinds of self-adaptive worms, and evaluate the performance of defensive schemes based on real-world traffic traces.