The Frontier Model Forum (FMF) released Version 1.0 of its Jailbreak Rating Framework for Frontier Cyber Threats on November 5, 2026. This system provides a standardized method for AI labs, security researchers, and policymakers to evaluate the risk of cyber jailbreaks, which occur when users bypass model safety guardrails to facilitate cyberattacks.
The framework utilizes a scoring system based on four input criteria to map threats across two axes: severity and likelihood. The resulting risk classification ranges from negligible to critical. This structure allows organizations to prioritize urgent threats and ensures consistent communication regarding security vulnerabilities across the industry.
Developed by FMF members with input from the AI Security Institute (AISI), the framework aims to replace inconsistent internal jargon with a shared industry standard. The FMF, which includes companies such as Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI, intends to update the document iteratively based on feedback.
The current version is considered preliminary, and the organization has opened a public comment period that will remain active until December 21, 2026. This initiative follows a February 2026 information-sharing agreement between the labs, further formalizing how major AI developers handle intelligence regarding security threats.
