Evo-Guard: Self-Evolving GNN Guardrails for Adaptive Safety in GUI Agents

Autonomous GUI agents operating in dynamic environments face significant safety risks that are poorly addressed by static guardrails. We propose a self-evolving graph-based guardrail framework $\textbf{EVO-Guard}$ for GUI agents that models interaction trajectories as structured graphs and leverages a Graph Neural Network (GNN) memory to predict both execution risk and violated safety rules. An interpretable arbiter integrates these predictions to regulate agent actions at inference time. Moreover, the framework continuously abstracts new atomic rules from high-risk trajectories, enabling adaptive safety reasoning without manual reprogramming. Experiments show improved safety prediction accuracy and generalization over static and non-graph-based baselines.