The Frontier Model Forum makes a risk scale for cyber jailbreaks

On 5 November 2026, the Frontier Model Forum released Version 1.0 of a framework to rate cyber jailbreaks. The Forum is a group of AI labs. A jailbreak is a trick that makes an AI model ignore its safety rules. The scale shows how bad a jailbreak is when it works. The Forum tells that this is a first version, and it can change.

How it works

The framework rates a jailbreak on four inputs:

• New function: how much new attack function the trick gives.
• Width: how wide that function is.
• Use: how easily a person can use the trick.
• Find: how easily a person can find the trick.

The four inputs give a Severity score and a Likelihood score. These give one class, from low to top.

The numbers

The Forum gave no test results. Persons at the AI Security Institute gave input. They tell that the framework is a good step.

What is not known

Labs in the Forum made the framework, and their models are the models that it rates. Only the institute gave external input. The Forum will change the framework after comments, which close on 21 December 2026.

Insight: The labs write the scale that rates them

The Forum made the scale, and only the institute gave external input. In a test of 9 top AI models, the risk rate went from 21.7% to 54.5% when researchers added pressure. This reads as a race to set the measure: the group that writes the scale can possibly set the terms that regulators use then. Persons who use AI models can tell vendors to show their classes.