The history of the digital age demonstrates that policymakers often struggle to assign responsibility for security and safety until long after new technologies have been adopted, putting individuals and society at risk. This project aims to provide a roadmap for policymakers to work together with frontier AI organisations to secure advanced AI systems from cyber theft.

As advanced AI systems become more capable, there is an increased risk that these systems will be subject to theft, manipulation, and abuse by malicious actors. The RAND Corporation has proposed security level benchmarks to guide frontier AI organisations in protecting their advanced systems from potential threats. Yet implementing these benchmarks presents key challenges. How can the benchmark security measures be achieved in practice? What actions should governments take to translate the security measures into voluntary commitments and emerging AI regulatory regimes?

Guided by these questions, the project will:

  • Estimate the quantitative cost of interventions for achieving the security level benchmarks detailed in the RAND study;
  • Examine incentive structures and institutional practices within AI companies that promote or inhibit the ability of achieving the security levels; and
  • Determine what level and form of government intervention is required.

The project will produce a final report offering a blueprint for securing advanced AI systems from cyber theft. Bringing together leading experts in AI and cybersecurity, the project seeks to establish a model for public-private collaboration and drive innovative approaches to learning and governance on AI security.