More details on Fable 5’s cyber safeguards and our jailbreak framework

2026-07-02 · Source: Anthropic News · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, extended

Summary

Claude Fable 5, now globally available, incorporates enhanced cybersecurity safeguards and introduces a proposed AI jailbreak severity framework. Its safety classifiers categorize cybersecurity activities into four types: Prohibited, High-risk dual use, Low-risk dual use, and Benign use, with a "safety margin" designed to prevent harmful outputs by cautiously blocking some benign or low-risk requests. Concurrently, Anthropic, in collaboration with Glasswing partners, has drafted a Cyber Jailbreak Severity (CJS) framework. This framework grades AI jailbreaks from CJS-0 (Informational) to CJS-4 (Critical) based on four axes: Capability gain, Breadth of capability gain, Ease of weaponization, and Discoverability. This initiative aims to standardize the assessment and communication of AI jailbreak risks among AI developers, governments, and the broader community.

Key takeaway

For AI Security Engineers evaluating model risks or Policy Makers developing AI safety standards, you should familiarize yourself with Claude Fable 5's cybersecurity classifier categories and the proposed Cyber Jailbreak Severity (CJS) framework. This framework offers a standardized method to assess jailbreak impact across capability gain, breadth, weaponization ease, and discoverability. Utilize these definitions to consistently classify and communicate AI security risks, contributing to a common language for safer AI deployment and more effective defensive strategies.

Key insights

The article details Fable 5's dual-use cybersecurity safeguards and a four-axis framework for grading AI jailbreak severity.

Principles

• Cybersecurity capabilities are often dual-use.
• Jailbreak severity depends on attacker uplift.
• A safety margin reduces harmful outcomes.

Method

The Cyber Jailbreak Severity (CJS) framework assesses jailbreaks by summing scores from four axes: Capability gain, Breadth of capability gain, Ease of weaponization, and Discoverability, yielding a CJS-0 to CJS-4 level.

In practice

• Submit cyber jailbreaks to HackerOne program.
• Provide feedback on the CJS framework.
• Classify cybersecurity uses into four categories.

Topics

• AI Safety
• Cybersecurity Classifiers
• AI Jailbreak Framework
• Claude Fable 5
• Dual-use AI
• Vulnerability Management

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Security Engineer, AI Scientist, Policy Maker

Related on AIssential

• Mythos BANNED (explained) — All AI models are inherently jailbreakable, as demonstrated by Fable 5's rapid bypass and subsequent shutdown, underscoring universal vulnerability.
• Quoting Matteo Wong, The Atlantic — Fable's refusal to "review" but compliance to "fix" insecure code is deemed intended cyberdefense functionality.
• Anthropic's Fable 5 is back worldwide after a two-week government ban over a jailbreak — AI models, even with enhanced guardrails, remain inherently vulnerable to jailbreaks, necessitating continuous security evolution and industry standards.
• When a Government Pulls an AI Model: What the Fable 5 and Mythos 5 Suspension Means for Security Teams — Government-mandated suspension of AI models highlights critical reliability and security implications for dependent systems.
• Meet the AI jailbreakers: ‘I see the worst things humanity has produced’ — AI jailbreaking, while emotionally taxing, is critical for identifying and mitigating safety vulnerabilities in large language models.
• American Government Takes Down Claude Fable — Abrupt government AI export controls based on minor jailbreaks reveal regulatory immaturity and risk hindering US AI leadership.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Anthropic News.