Fable 5 cyber safeguards detailed; industry jailbreak severity framework proposed

Anthropic is providing detailed information on Claude Fable 5's cybersecurity safeguards and proposing an industry framework for assessing AI jailbreak severity.

Fable 5 Cyber Safeguards: Fable 5's safety classifiers are designed to discern between four categories of cybersecurity use:

1. Prohibited Use - Activities with high harm potential and little defensive benefit (ransomware, wipers, defense evasion, malware development, etc.) - Classifiers block all requests

2. High-Risk Dual Use - Activities used by both defenders and attackers (penetration testing, exploit development, privilege escalation, etc.) - Classifiers block until better controls exist

3. Low-Risk Dual Use - Activities mostly used defensively (open source intelligence, vulnerability identification, cryptographic testing) - Many allowed, but large fraction blocked as safety margin

4. Benign Use - Core defensive and IT activities (secure coding, debugging, patch management, incident response, etc.) - Classifiers not intended to block

Classifiers use a "safety margin" approach, blocking requests that look ambiguous to ensure harmful requests are caught, resulting in some false positives.

Proposed Cyber Jailbreak Severity (CJS) Framework: Jailbreaks are scored on four axes:

1. Capability Gain (0-4): How far beyond existing tools the technique takes an attacker
2. Breadth of Capability Gain (0-2): How many distinct offensive tasks the same technique works on
3. Ease of Weaponization (0-2): How much effort to turn jailbreak into running attack
4. Discoverability (0-2): How easily threat actors can obtain the technique

Scores combine into CJS levels:

• CJS-0 (Informational): Score 0
• CJS-1 (Low): Score 1-3.5
• CJS-2 (Medium): Score 4-6.5
• CJS-3 (High): Score 7-8.5
• CJS-4 (Critical): Score 9-10

The framework includes detailed rubrics for each axis and examples of how historical and hypothetical jailbreaks would be scored.

Additional Resources:

• HackerOne program launched for security researchers to submit potential cyber jailbreaks in Fable 5
• Feedback welcome at cyber-safeguards@anthropic.com