FMF Jailbreak Rating Framework for Frontier Cyber Threats (www.frontiermodelforum.org)

The Frontier Model Forum (FMF) has launched Version 1.0 of its Jailbreak Rating Framework, providing a systematic method to evaluate the risks associated with cyber jailbreaks in AI models. This initiative is particularly significant as it aims to bridge a critical gap in the understanding of how various jailbreak techniques can be exploited, consequently aiding AI developers, security experts, and policymakers to mitigate potential threats more effectively. With frontier AI models increasingly capable of both bolstering and undermining security, the framework enables organizations to assess and address jailbreaks consistently based on their risk profiles. The framework employs a tiered assessment approach that evaluates jailbreaks across four categories: capability gain, capability scope, ease of deployment, and ease of discovery. This structured methodology not only promotes comparability in assessments but also helps prioritize responses to high-risk jailbreaks over low-risk ones. Key industry figures from companies like Amazon, Google DeepMind, Anthropic, Microsoft, and OpenAI have expressed their support, emphasizing the shared need for effective safeguards as AI technologies advance. The FMF intends to refine the framework further through continuous testing and public feedback, highlighting its commitment to enhancing AI security in a collaborative manner.