Oxford Martin AIGI report maps open problems in frontier AI risk management standards

A paper dated February 24, 2026, written by a large team of twenty-nine named authors that includes Marta Ziosi, Stephen Casper, Henry Papadatos and Robert Trager, sets out unresolved problems in the way risks from frontier AI are managed. Frontier systems are general-purpose models able to carry out a wide range of tasks, and the paper starts from the observation that they create safety risks which risk management can help address.

The authors explain why this is difficult. Most risk management standards written for AI were designed for narrow systems and predate frontier models, which both magnify familiar risks and bring qualitatively new ones. The rapid pace of change has left no stable scientific consensus, and new frontier safety practices are frequently out of step with established risk management frameworks or may even weaken them.

To respond, the paper takes a problem-oriented approach. Through a structured literature review it walks through each stage of risk management, namely planning, identification, analysis, evaluation and mitigation, and records the open problems at each stage along with who is best placed to tackle them. Problems are sorted into three groups: those reflecting missing scientific or technical consensus, those reflecting conflict with existing frameworks, and those where implementation falls short even though agreement and alignment appear to exist. The actors it names are developers and deployers, regulators, standards bodies, researchers and outside evaluators.

The paper states that it offers no particular solutions. Instead it is meant as an agenda-setting reference, accompanied by a living online repository, to help coordination, limit duplicated effort and steer future research and governance work.