Anthropic proposes a severity framework for AI jailbreaks

Anthropic published a draft framework for describing AI jailbreak severity on July 2, alongside more detail about Claude Fable 5’s cyber safeguards. The technical post argues that a shared vocabulary would help developers, governments, researchers, and customers distinguish a narrow prompt bypass from a broadly reusable route to dangerous capability.

A bypass needs impact dimensions

The draft considers what new capability a jailbreak unlocks, how broadly it works across targets or techniques, and how reliably it can be reproduced. Anthropic illustrates breadth with a scale ranging from one question or artifact to techniques that generalize across vulnerability classes and targets. The post also describes Fable 5 classifier categories: prohibited use, high-risk dual use, low-risk dual use, and benign use, with different intended blocking or monitoring behavior.

This is more actionable than counting successful adversarial prompts. A security report should identify the model and safeguard version, exact prerequisites, number of attempts, eligible target set, harmful action enabled, and whether minor paraphrases preserve the bypass. Without those fields, two reports called “jailbreaks” can represent very different operational risks.

Draft scores still need governance

Anthropic presents the framework as current thinking and requests critique. It also opened a HackerOne channel for Fable 5 cyber-jailbreak submissions. A scoring system can improve triage, but owners must define who verifies reproduction, how embargoes work, when model or classifier changes invalidate a score, and what customer notification each level triggers.

Teams should adapt the dimensions into their vulnerability intake form now, while keeping the draft label visible. Test the rubric against several past reports and check whether independent reviewers reach similar ratings. Disagreement is useful: it exposes terms that need clearer evidence before the score can drive release or disclosure decisions.