Source: Anthropic
What was announced
Anthropic has re-deployed Claude Fable 5 globally and provided more details on its cybersecurity safeguards and a proposed AI jailbreak severity framework.
Why it matters
Developers can now understand what behaviors are blocked by Fable 5's safety classifiers, and they have access to a draft framework for assessing the severity of AI jailbreaks. This information helps in making informed decisions about model usage and security.
Key takeaways
- Fable 5's safety classifiers now have a detailed list of blocked and non-blocked behaviors.
- Developers can use the proposed AI jailbreak severity framework to evaluate and discuss the risks posed by AI jailbreaks.
- Developers should review the framework and provide feedback to Anthropic at cyber-safeguards@anthropic.com to help shape future guidelines.