AI Cyber Jailbreak Risk 5-Level Assessment Framework Released… Public Comments Accepted Until December 21

The Frontier Model Forum released version 1.0 of a five-tier framework for rating the severity of AI cyber jailbreaks, open for public comment until December 21. The rubric scores individual, reproducible cases in which a specific version of a frontier model bypasses its safeguards, covering multi-turn conversations and automated tooling rather than single prompts alone. The FMF said the common scale is meant to let developers, security researchers and policymakers compare incidents and set response priorities. It matters because the field has lacked a shared vocabulary for ranking jailbreak severity as AI systems face growing security scrutiny. What to watch next is whether member labs and outside researchers adopt the tiers during the comment period and whether the final version adds binding thresholds for disclosure.