Details on Fable 5's Cybersecurity Safeguards and Jailbreak Framework

Alongside the global rollout of Claude Fable 5, Anthropic shared details of the Cybersecurity Safeguards and the AI Jailbreak Severity Framework designed to prevent misuse of the model.

Cybersecurity techniques have a dual-use nature, serving both defensive and offensive purposes. Accordingly, rather than blocking all security-related activity, Anthropic manages it by classifying it into four risk-based categories.

• Prohibited use: Activities with little defensive value that can cause serious harm (blocked)
• High-risk dual use: Activities widely used by malicious actors but that also have beneficial applications (blocked)
• Low-risk dual use: Activities used mainly for defensive purposes but with potential for misuse (monitored and blocked when necessary)
• Benign use: Activities that cause no harm (allowed and monitored)

Anthropic also presented a draft framework for defining the severity of jailbreaks, which bypass AI safeguards. This aims to enable consistent risk communication between developers and governments, and is paired with a HackerOne program through which security researchers can submit discovered jailbreaks.