Addressing risks not closely tied to dangerous capabilities (systemic risks and AI malfunction)

Characterize and integrate into frontier AI safety cases risk domains that are not closely tied to specific dangerous capabilities, including systemic risks and risks from AI malfunction, and develop assessment and mitigation approaches appropriate for those risks within the safety case structure.

Background

While current safety case approaches emphasize dangerous capabilities, the authors identify open questions about how to address risks that are less closely tied to specific capabilities, such as systemic risks or risks arising from AI malfunction. These risk types require alternative methods of assessment and justification to be coherently represented in safety case arguments.