Author: Ed Pocock
Anthropic's release of Mythos under Project Glasswing1 has created a fourth category of frontier AI model: restricted distribution. This forces the question: what are the implications for the reinsurance industry when AI's most capable models are deliberately kept out of the hands of those we rely on to evaluate them?
Frontier AI models have typically been open-source, open-weight or fully proprietary, but Anthropic's Mythos takes a different approach, with restricted access due to its reported performance to identify vulnerabilities and generate exploits in software.2 This reflects real concern about misuse, with past incidents showing how quickly advanced capabilities can fuel widespread cyberattacks.
Building on the example of Mythos, this evaluation explores the wider implications of restricted AI models for insurers and extends the insights from our Q1 InsurTech report on AI risk. To assess and underwrite this evolving risk effectively, insurers need to understand how AI models behave in practice. Robust model evaluation will be a critical foundation for this emerging market, but current methods for benchmarking AI performance will need to evolve significantly to keep pace.
To understand what this means for risk, resilience and the future threat landscape, read the full whitepaper.
