Jailbreak Rating Framework

The Jailbreak Rating Framework is a Frontier Model Forum (FMF) draft specification for prioritizing risk when an AI safeguard bypass enables serious cyber harm — not for estimating absolute damage.

Version 1.0 scores successful jailbreaks on capability gain, capability scope, ease of deployment, and ease of discovery, mapping to risk bands from Negligible to Critical. OpenAI and other member labs helped drive it.

Update: https://www.frontiermodelforum.org/updates/introducing-the-fmf-jailbreak-rating-framework-for-frontier-cyber-threats/ Spec: https://www.frontiermodelforum.org/specifications/jailbreak-rating-framework-for-frontier-cyber-threats/

Why it matters

Labs lacked a shared language for jailbreak severity. The framework aims at a common rating method for prioritization and cross-org communication.

Caveats

• Preliminary; feedback open through 2026-12-21;
• Scoped to frontier cyber-relevant jailbreaks — not a universal content-safety ruler;
• Name ≠ exploit how-tos — this entry is the FMF rating spec itself.

1. Read the FMF update and v1.0 specification;
2. Score your red-team jailbreak finds with its dimensions;
3. Submit feedback to FMF during the comment window;
4. Align with internal security process — not a legal liability standard.