The Frontier Model Forum publishes a framework for rating cyber jailbreaks
On 6 November 2026 OpenAI announced a Frontier Model Forum framework for comparing cyber jailbreaks. This note does not describe a technique. Public comments run through 21 December 2026.
On 6 November 2026 @OpenAI announced a partnership with the Frontier Model Forum and member firms. The object is version 1.0 of the Jailbreak Rating Framework for Frontier Cyber Threats. The forum page describes a preliminary method for comparing the risk profile of cyber jailbreaks that have already succeeded, so different teams do not spend effort first on the lower-risk case.
The page names four rating categories: capability gain, capability scope, ease of deployment, and ease of discovery. It does not reproduce the scale, the operational example, or the matrix. That stays out on purpose. A note from this desk does not teach how to bypass a safeguard.
The forum page extract did not return a publication date. The date used here is the OpenAI post, 6 November. Public comments run through 21 December 2026, to [email protected]. The page says version 1.0 will still be tested, to see whether different assessors reach similar ratings.
Sasha Baker, OpenAI’s head of national security policy, is quoted on the page and in the post. The page also quotes Hart Rossman of Amazon Web Services, Sarah Heck of Anthropic, Owen Larter of Google DeepMind, and Natasha Crampton of Microsoft. When this desk read the post, X showed 344,300 views.