Open Problems in Frontier AI Risk Management

A large team of 29 researchers led by Marta Ziosi has published a comprehensive 81-page paper on arXiv titled 'Open Problems in Frontier AI Risk Management,' which systematically maps unresolved challenges across the entire risk management lifecycle. The paper examines each stage—risk planning, identification, analysis, evaluation, and mitigation—through a structured literature review, surfacing open problems that currently hinder robust governance of frontier AI systems.

Critically, the authors classify problems into three distinct categories: (a) lack of scientific or technical consensus, such as disagreements over evaluation benchmarks or failure modes; (b) misalignment with, or challenges to, established risk management frameworks, where emerging AI safety practices conflict with existing regulatory approaches; and (c) shortcomings in implementation despite apparent consensus, where known best practices are not effectively applied. The paper also identifies the actors best positioned to address each problem—including developers, deployers, regulators, standards bodies, researchers, and third-party evaluators—and provides a living online repository to support ongoing coordination and reduce duplication of effort across the field.