All news
AI, LLM & Agents·10 days ago

Anthropic proposes a Cyber Jailbreak Severity scale (CJS 0–4) with Amazon, Microsoft, Google

AnthropicSummarized by the ORA·tech AI assistant
AI

Anthropic detailed Claude Fable 5's safety classifiers and proposed a Cyber Jailbreak Severity (CJS) framework that scores a jailbreak's danger on a 0–4 scale. For developers building on AI, it offers a shared way to rate how serious a successful jailbreak really is, rather than vendor-by-vendor safety claims.

With Claude Fable 5 redeployed globally, Anthropic published more detail on the model's cybersecurity safeguards and proposed a shared industry framework, developed with Amazon, Microsoft and Google under Project Glasswing.

Key points

  • A four-tier classifier in Fable 5 blocks prohibited uses (ransomware, wipers, cyber-physical sabotage, defense evasion, malware development/delivery, C2 infrastructure, backbone attacks) while allowing legitimate defensive work (secure coding, patch management, SOC analysis, malware reverse-engineering, incident response) with minimal restrictions.
  • The Cyber Jailbreak Severity (CJS) scale runs 0–4, scored on four measures: capability gain, breadth of that gain, ease of weaponisation, and discoverability.
  • Summed scores map to bands: CJS-1 Low (1–3.5), CJS-2 Medium (4–6.5), CJS-3 High (7–8.5), CJS-4 Critical (9–10).
  • It is a proposed, open framework meant to standardize how vendors and regulators rate the severity of a successful jailbreak.
Source
Anthropic
Read the original
#Anthropic#Claude Fable 5#AI safety#jailbreak#cybersecurity
This summary was written by the ORA·tech AI assistant. Read the original for full context.

Related

Chào bạn, mình là Nova. Có phần nào bạn muốn hiểu kỹ hơn không? Cứ nói với mình nhé.