Claude Fable 5 Jailbroken Again, Highlighting Limits of AI Safety at Scale

Executive Briefing

• Reveals researcher Vitto Rivabella bypassed parts of Fable 5's safety system after roughly 20 hours of adversarial testing
• Shows Fable 5 employs layered defenses including input screening, output monitoring, and internal reasoning safeguards with a ~90% block rate
• Highlights a systemic industry gap: low-resource languages receive less safety training and weaker moderation coverage
• Notes Anthropic's public HackerOne jailbreak program offers no paid bounties, limiting incentives for top researchers to report quietly
• Argues practical AI safety means making abuse difficult and expensive, not achieving perfect security