Anthropic Proposes Five-Tier Industry Standard for Rating AI Jailbreak Severity

Summary: Anthropic published the Cyber Jailbreak Severity (CJS) framework — a five-tier, exponential scale co-developed with Amazon, Microsoft, and Google to standardize how dangerous an AI jailbreak is rated.

Key Points

• Five tiers from CJS-0 (Informational) to CJS-4 (Critical); each step represents exponentially greater risk than the last
• Severity is scored across four axes: capability gain, breadth of harm, ease of weaponization, and discoverability
• Scope is limited to cybersecurity jailbreaks; non-cyber misuse (e.g., system-prompt extraction) is explicitly excluded
• Released alongside a new Claude Fable 5 safety classifier that Anthropic says blocks over 99% of attempts replicating the June jailbreak that triggered a 19-day US export-control suspension

Why It Matters

Until now, there has been no shared vocabulary for comparing jailbreak incidents across the AI industry. If adopted, CJS gives companies, regulators, and security researchers a common scale — analogous to CVSS scores in traditional cybersecurity — to communicate incident severity and set consistent remediation thresholds.

Read More

• Anthropic official post — Anthropic
• Deep dive analysis — CybersecurityNews