Skip to content

AI security

Frontier Model Forum rates cyber jailbreaks on four scores

2 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: OpenAI models and products →
Image: x.com

The Frontier Model Forum on November 5 put out a common scale for how dangerous a successful cyber jailbreak is. It scores each one on four counts: capability gain, capability scope, ease of deployment and ease of discovery. If your team triages jailbreak reports, treat it as unvalidated for now: this is a preliminary Version 1.0, the forum has yet to test whether different assessors reach similar ratings, and it is taking outside feedback through December 21, 2026.

The Frontier Model Forum’s announcement defines a jailbreak as a reproducible technique that unlocks capabilities or information a model’s safeguards are meant to restrict. Some produce responses of limited practical use. Others could give attackers dangerous cyber capabilities. Without a shared rating, the forum argues, organizations score the same jailbreak differently, struggle to compare findings and spend limited resources patching lower-risk issues before high-risk ones.

The method is tiered and combines the scores from all four categories. Capability gain and scope describe what a jailbreak unlocks. Ease of deployment and ease of discovery describe how readily someone can find it and put it to use. The point is to make assessments of the same bypass more consistent from one team to the next.

OpenAI, Anthropic, Google DeepMind, Microsoft and Amazon all put out statements backing the release. Anthropic, Google, Microsoft and OpenAI founded the forum as an industry body on July 26, 2023, so the scale comes from the companies whose safeguards it rates. OpenAI says it worked with the forum and its member firms to help identify when bypassing a model’s safeguards could enable serious cyber harm. “We’ve long known that guardrail bypasses vary widely in their severity, but until now there hasn’t been an agreed-upon methodology for assessing their impact,” said Hart Rossman, VP of Security at Amazon Web Services.

The forum plans to keep running the framework against a range of jailbreaks. It wants to check two things: whether assessors agree, and whether the ratings reflect real differences in risk. It also says it will revise the framework as new jailbreaks turn up and existing techniques evolve. The framework belongs to the forum’s wider AI-cyber workstream, which also covers methods for assessing frontier cyber capabilities of concern.

Feedback goes to comments@frontiermodelforum.org. The forum says it will review submissions with its members and work relevant input into an updated version.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.