Anthropic Details Claude Fable 5 Cyber Safeguards and AI Jailbreak Severity Framework

Anthropic published technical details on Claude Fable 5's cybersecurity safeguards and an early-draft five-band AI jailbreak severity framework: classifiers sort requests into four tiers (prohibited, high-risk dual use, low-risk dual use, benign) with a wider safety margin, and a combined-score severity model (capability gain, breadth, ease of weaponization, discoverability) sets a floor that can be raised for discretionary factors; Anthropic is soliciting feedback and running a HackerOne bounty for reported cyber-jailbreaks.