New framework grades AI jailbreaks across capability gains, scope, and how easy they are to deploy
Reported by Tom's Hardware1 source ↓
The Frontier Model Forum released a draft framework for rating AI jailbreaks by the capabilities they add, their scope, and how easy they are to find and use. Comments are open until December 21.
Key points
- The framework rates capability gain, capability scope, ease of deployment and ease of discovery from 0 to 3
- Its combined rating ranges from low to critical, based on severity and likelihood
- Version 1.0 is a work in progress, and the forum advises against relying on it for now
- The comment period runs until December 21
Details
The Frontier Model Forum’s framework is intended to classify jailbreaks that enable cyber capabilities beyond those already available through other models or tools. The forum says capability gain is the central measure and sets limits on the other scores; a jailbreak that adds no new capability receives a zero on that measure.
The four component scores cover capability gain, scope, deployment and discovery. The framework combines them under severity and likelihood to produce an overall rating from low to critical. The forum says the baseline will shift as models and alternative tools change, and that a zero does not mean a flaw should be ignored.
Written from coverage as of
Images
- Tom's Hardware