CVSS-Style Rating System for AI Models from Frontier Model Forum

The Frontier Model Forum (FMF) has announced its new Jailbreak Rating Framework to monitor how easily artificial intelligence models can bypass security restrictions and to estimate the scope of potential cyber harms.

The new framework published by the Frontier Model Forum operates with an approach similar to the CVSS system commonly used in grading security vulnerabilities. The system aims to classify the types of jailbreaks to which large language models (LLMs) are exposed and their severity across various axes.

Established in 2023 with the support of Anthropic, Google, Microsoft, and OpenAI, the project also focuses on estimating the scale of potential harms of artificial intelligence models that could create potential cyber risks across the internet.