This thesis proposes a formal methodology for defining, specifying, and reasoning about micro-policies — security policies based on fine-grained tagging that include forms of access control, memory safety, compartmentalization, and information-flow control. Our methodology is based on a symbolic machine that extends a conventional RISC-like architecture with tags. Tags express security properties of parts of the program state ("this is an instruction," "this is secret," etc.), and are checked and propagated on every instruction according to flexible user-supplied rules. We apply this methodology to two widely studied policies, information-flow control and heap memory safety, implementing them with the symbolic machine and formally characterizing their security guarantees: for information-flow control, we prove a classic notion of termination-insensitive noninterference; for memory safety, a novel property that protects memory regions that a program cannot validly reach through the pointers it possesses — which, we believe, provides a useful criterion for evaluating and comparing different flavors of memory safety. We show how the symbolic machine can be realized with a more practical processor design, where a software monitor takes advantage of a hardware cache to speed up its execution while protecting itself from potentially malicious user-level code. Our development has been formalized and verified in the Coq proof assistant, attesting that our methodology can provide rigorous security guarantees.