Anthropic Details Claude Fable 5 Cyber Safeguards and Jailbreak Framework

Anthropic shared further details on Claude Fable 5, which has been re-deployed globally for all users. The update covers the cybersecurity safeguards—specifically safety classifiers—that accompany the model, and an early draft of an AI jailbreak severity framework developed with Glasswing partners.

The safety classifiers categorize cybersecurity use into four types: prohibited use, high-risk dual use, low-risk dual use, and benign use. Prohibited use includes activities such as ransomware, wipers, cyber-physical sabotage, defense evasion, command-and-control, data exfiltration, malware development, and internet backbone attacks, which the classifiers are intended to block. High-risk dual use covers actions like penetration testing, unauthorized access, privilege escalation, exploit development, and high-uplift vulnerability finding; these are blocked by default until better controls limit access to known good actors. Anthropic noted that it does not intend to block all vulnerability finding, but aims to prevent the model from identifying vulnerabilities that other widely available models cannot.

The proposed jailbreak framework is intended to standardize descriptions of jailbreak severity so that AI developers and governments can discuss risks in consistent terms. Anthropic also launched a HackerOne program for researchers to submit potential cyber jailbreaks in Fable 5 and invited feedback at cyber-safeguards@anthropic.com.

Source: Anthropic News