The rapid development of AI, particularly agentic systems, has outpaced scientific understanding of its risks, safety, and societal impacts. Current testing methods, including benchmarks and red-teaming, often overlook human behavior, institutional contexts, and fundamental limits of testing, challenging their ability to anticipate harms and support trustworthy and safe deployment. This workshop brings together researchers from academia, non-profit, and industry, and policymakers to strengthen the empirical and theoretical foundations of AI testing, evaluation, and governance. Participants will examine the limitations of existing approaches, broaden the scope of assessment, and identify opportunities for interdisciplinary collaboration. The workshop aims to develop a shared vocabulary for describing evolving AI risk, evaluation and governance across different levels of severity and time horizons, alongside an adaptive evaluation framework suited to increasingly autonomous systems. Its central goal is to advance a rigorous, inclusive science of AI risk assessment, mitigation and governance.
About
Chairs/Organizers