UK AI Security Institute: Safety Cases
Source
Safety Case Template for Frontier AI: A Cyber Inability Argument
The UK AI Security Institute's research develops a template for constructing an evidence-based argument that an AI system lacks capabilities sufficient to pose unacceptable offensive cyber risk.
The template organizes risk models, proxy tasks, evaluation settings, evaluation results, and supporting arguments using a Claims, Arguments, Evidence structure.
Pilot Investigation
The Structural Assurability investigation examines which parts of a cyber-inability argument concern the availability of evidence capable of distinguishing claim-relevant alternatives.
It also considers which relationships require additional justification, including those connecting proxy-task performance to risk-relevant capabilities.
The investigation distinguishes limitations expressible through the current formal theory from obligations involving measurement validity, inference, and the broader safety argument.
Research Status
AISI-001 remains OPEN.
An initial conceptual crosswalk has been developed. Detailed source review, selection of a specific risk model and proxy task, and construction of an applicable observation model remain outstanding.