Efficiency at Risk: Entropy-Induced Failure in Learned Decision Policies under Partial Observability

This work presents an empirical evaluation of automated security probing policies operating under partial observability. We quantify efficiency gains achieved by learned decision models and identify entropy-induced failure regimes that emerge under increasing observational noise. By modeling the interaction as a Partially Observable Markov Decision Process (POMDP), we demonstrate that while learned policies significantly reduce average Time-to-Compromise, they exhibit heavy-tailed operational risk characterized by instability, action thrashing, and elevated detection likelihood. We further propose a hybrid meta-control architecture that bounds worst-case behavior by supervising learned policies with entropy-aware fallback mechanisms. Our results highlight a fundamental efficiency–stability trade-off and provide actionable guidance for the safe deployment of autonomous security decision systems in stochastic environments. -