OpenAI and Frontier Model Forum Ship Jailbreak Rating Framework v1.0 for Frontier Cyber Threats

Core Background & Industry Pain Points

As frontier models gain cyber capability, not every successful jailbreak is equal: some yield low-value text, others may unlock actionable attack capability. Without a shared severity language, teams can burn cycles on lower-risk bypasses. On 2026-11-05 the Frontier Model Forum released the Jailbreak Rating Framework for Frontier Cyber Threats v1.0; on 2026-11-06 OpenAI amplified it with a quote from Head of National Security Policy Sasha Baker.

Architecture Highlights & Internals

The framework uses a tiered assessment across four scores: capability gain versus baseline, capability scope, ease of deployment, and ease of discovery. The intent is to compare jailbreaks by what they unlock and how readily those capabilities can be used—not merely whether a bypass succeeds. FMF will keep testing assessor agreement and iterate as techniques evolve.

Authoritative Benchmarks & Measured Scores

This release is a shared methodology, not a public leaderboard or single cutoff score. OpenAI, Amazon, Anthropic, Google DeepMind, and Microsoft provided supporting quotes on the FMF page. Public comments run through 2026-12-21. It does not replace vendor red teams or CVE processes.

Developer Hands-on Guide

For agent/tool products, triage successful jailbreak tickets with the four axes so every bypass is not treated as equal severity. Red-team writeups should state capability gain and deployment cost. Track the FMF comment window and later revisions; production defenses still rest on vendor safety layers plus your own refusal policy.