What happened
The Frontier Model Forum published a new shared framework (Version 1.0, open for public feedback until 21 December 2026) for rating successful jailbreaks against frontier models. It notes that "there is no shared approach to making these assessments, which makes it harder to compare findings and identify the kinds of jailbreaks that demand more attention." The framework scores a single jailbreak across four categories — capability gain, capability scope, ease of deployment, and ease of discovery — using logarithmic ordinal tiers defined both conceptually (plain language) and operationally (concrete expert tests). It is scoped to "frontier risk," deliberately measures relative rather than absolute harm, and is designed to make findings comparable across teams and organizations.
Why it matters
It establishes the first vendor-neutral measurement language for agentic/cyber jailbreak risk, which front-line security and safety teams can adopt now and which will shape how frontier-cyber incidents are communicated to boards and regulators.
Action needed
Map the FMF jailbreak rating framework onto your AI-security evaluation program and submit feedback before the public-comment window closes in December.