Frontier labs publish a shared scale for cyber jailbreaks

On Nov. 5, 2026 the Frontier Model Forum released version 1.0 of a framework that rates one successful cyber jailbreak on capability gain, capability scope, ease of deployment, and ease of discovery, then maps those scores to bands from Negligible to Critical. Amazon, Anthropic, Google DeepMind, Microsoft, and OpenAI each said they contributed. The forum says the tiers are a starting point rather than operational guidance, comments are open through Dec. 21, and feedback from NIST, the NSA, and UK cyber bodies does not imply endorsement.