Challenges of Restricted AI Models for Insurers

This document is a technical report discussing the implications of restricted AI models, particularly Anthropic's Mythos, on the insurance industry. It outlines the categorization of large language models into open-source, open weight, and proprietary models, introducing a new category of restricted distribution. The report explains how restricted access to advanced AI models can impact the risk assessment and pricing strategies of insurers. It highlights historical precedents, such as the 2017 cyber insurance market hardening due to the release of NSA exploits, and discusses the urgency for insurers to understand AI behavior to manage risks effectively. The report emphasizes the importance of independent model evaluations to provide objective assessments of AI capabilities and the potential consequences of withholding models from evaluators. It concludes with a call for improved evaluation techniques to ensure that insurers can accurately price AI-related risks in an evolving landscape.