← All stories
● Covered by 1 source · 1 reportMedium impact1 neutral

Frontier Model Forum Releases Jailbreak Rating Framework for AI Models

🔄 Updated 11d ago
New to BrevFeed? We gather this story from every outlet covering it into one summary — ranked by real-world impact, not just the latest headline — so you never miss what matters. What is BrevFeed? →

Key points

  • FMF released a new Jailbreak Rating Framework for AI models.
  • Framework classifies jailbreaks by severity and likelihood of weaponization.
  • It includes scores for capability gains, scope, deployment, and discovery ease.
  • The system is a work-in-progress, open for comments until December 21.

New Framework for AI Jailbreaks

The Frontier Model Forum (FMF), backed by Anthropic, Google, Microsoft, and OpenAI, has published its Jailbreak Rating Framework. This new system is designed to assess how easily AI models can circumvent their built-in safeguards and the potential extent of damage if such a breach occurs. It functions as an analog to the CVSS security rating system but is specifically tailored for Large Language Model (LLM) jailbreaks across various providers.

Standardizing Jailbreak Evaluation

While many AI companies currently use their own internal rating systems, the FMF's goal is to create a standardized, cross-vendor classification method. This standard is intended to provide a common reference for teams developing patches and countermeasures against AI model vulnerabilities. The framework is currently in version 1.0 and is open for public comments and suggestions until December 21.

Scoring Components and Categories

The framework consists of four scoring components organized into two main categories: severity and likelihood. Severity measures the seriousness of a jailbreak technique, while likelihood assesses how easy it is to discover and weaponize. These scores are combined to produce a final rating ranging from low to critical.

Individual ratings are assigned for capability gains, capability scope, ease of deployment, and ease of discovery, each graded from zero to three. The FMF emphasizes that 'capability gain' is the most critical dimension, as jailbreaks offering no new capabilities are considered less significant within the system.

Evolving Baseline

An important aspect of the FMF's framework is its dynamic baseline, which adjusts as AI models evolve. This means that a jailbreak rated as zero in the future may not imply the same level of harmlessness as it would today, reflecting the continuous advancement of AI capabilities and potential risks.

✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →

The daily brief

One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.

One email a day. Unsubscribe in one click, any time.

Today's brief

Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.

~4 min · 3 stories · Nov 21

▶ Play today's brief Listen on Spotify

New every morning, and the back catalogue is archived by date.

Reporting from

The Frontier Model Forum (FMF) has introduced a new Jailbreak Rating Framework to classify the severity and ease of deployment of AI model jailbreaks. This framework, supported by Anthropic, Google, Microsoft, and OpenAI, aims to establish a cross-vendor standard for evaluating how AI models bypass safeguards and the potential impact of such breaches.