Agent State Transition Security

Modern LLM-Agent systems transform security-relevant state across interaction modes, reasoning, memory, skills, principals, runtimes, and lifecycle stages. We introduce Agent State Transition Security (ASTS), a transition-centered formal framework for analyzing security effects and end-to-end invariants across heterogeneous Agent-state transitions. The model separates operational state dynamics, protected-effect interpretation, and normative treatment. A transition specification combines state evolution with a conservative cross-domain effect interpretation and partial contract translation; constrained transition-local obligations, authenticated event lineage, semantic-dependency assumptions, and relational contracts are used where the declared policy requires them. The theory distinguishes exact from conservative enforcement, proves an end-to-end safety result for transported prohibitions under complete mediation, and isolates separate no-recovery and relational-projection gaps. A clean-room finite artifact passes 14 deterministic test groups covering the stated witnesses and counterexamples. These tests check the reference model within a finite scope; they do not estimate prevalence or establish universal behavior in deployed Agent systems.