Anthropic has detailed its new Cyber Jailbreak Severity (CJS) scale for its Fable 5 model, introducing a four-tiered system to classify and score jailbreak attempts. This scale, ranging from 0 to 4, evaluates the seriousness of discovered vulnerabilities rather than the user's prompt itself. The company has also expanded its safety measures, intentionally increasing false positives to ensure malicious requests are blocked, a trade-off developers should be aware of when working with Fable 5. AI
IMPACT Provides developers and security professionals with a clearer understanding of AI model vulnerabilities and Anthropic's approach to safety.
RANK_REASON Publication of a new security classification system and scale for an AI model. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
