Shielding is a prominent method in runtime verification that offers formal safety guarantees for autonomous agents. By using a safety model of the environment, shielding enables the automatic construction of a mechanism that blocks unsafe actions during both training and deployment. In this talk, we explore the main components of the shielding pipeline. We begin by revisiting its central assumption: the availability of a world model that accurately captures the environment. We then discuss the requirements such models must fulfil, along with recent progress in learning discrete world models. Next, we examine the synthesis of shields that provide quantitative safety guarantees in environments that inherit both probabilistic and adversarial behaviour. We conclude by discussing shield representations that support human-understandable explanations of the shield’s decisions.

Bettina Könighofer is an Assistant Professor of Formal Methods and Machine Learning at Graz University of Technology. Her research focuses on runtime assurance, probabilistic model checking, and reinforcement learning. Her work on shielding was among the first to provide formal safety guarantees for deep reinforcement learning, bootstrapping the field of shielded learning. She received her PhD from TU Graz under the supervision of Prof. Roderick Bloem in 2020. Before joining the faculty in 2023, she led the TrustedAI group at LAMARR Security Research.