More details on Fable 5’s cyber safeguards and our jailbreak framework

Claude Fable 5 is now available globally with updated cybersecurity safeguards. These safeguards use AI classifiers to monitor and block potentially dangerous requests while allowing benign activities. Because cybersecurity tools are dual-use, the system categorizes requests into four levels: prohibited, high-risk dual use, low-risk dual use, and benign. To ensure safety, the model employs a large safety margin, which intentionally blocks some benign or low-risk requests to prevent harmful ones from slipping through.

The release also introduces a draft framework for assessing the severity of AI jailbreaks, which are methods used to bypass these safety protections. This framework aims to create a standardized language for developers, governments, and researchers to discuss risks. It evaluates jailbreaks based on four criteria: the level of new capability provided to an attacker, the breadth of tasks the technique can perform, how easily the technique can be turned into a real-world attack, and how easily the technique can be discovered.

These severity scores range from informational to critical, helping to categorize the danger posed by specific bypass attempts. To support this effort, a new program on HackerOne allows security researchers to submit discovered jailbreaks for review. This initiative is intended to foster collaboration across industry and government to establish consistent safety standards that permit defensive cybersecurity work while preventing malicious misuse of the technology.