Anthropic proposes a CVSS-style scale for AI jailbreak severity
The Cyber Jailbreak Severity scale rates exploits across four axes, summing them into five bands from Informational to Critical.
MVMara Vidović
Published 8 Jul 2026, 14:44Updated 29 Jul 2026, 04:291 min read
- ▸Anthropic's proposed Cyber Jailbreak Severity (CJS) scale scores exploits on four axes: capability gain, breadth of impact, ease of weaponization and discoverability
- ▸Scores sum into one of five severity bands, from CJS-0 (Informational) to CJS-4 (Critical)
- ▸The model is explicitly based on the Common Vulnerability Scoring System used by security teams since 2005
- ▸Adoption across multiple labs would standardize how jailbreak risk gets communicated publicly
Borrowing from two decades of security practice
Modeling AI jailbreak severity on CVSS — rather than inventing a new framework from scratch — lets the industry reuse two decades of institutional familiarity among security teams instead of asking them to learn an entirely new risk language.
Source: Anthropic
