Anthropic Publishes Fable 5 Cyber Safeguards and a Jailbreak Severity Framework

On July 2, 2026, Anthropic published details of the safety classifiers protecting the cyber-capable Claude Fable 5 and proposed an early framework, the Cyber Jailbreak Severity (CJS) scale, for grading how serious a jailbreak of a model's safeguards is, in an official blog post. The company laid out how it aims to curb harmful use of models with dual-use cyber capabilities while preserving legitimate work.