Red Teaming Wasn't Enough: Claude Fable 5 Jailbroken in 48 Hours

Anthropic's red-team bounty program failed to prevent a sophisticated jailbreak of Claude Fable 5, exposing a fundamental weakness in model-layer guardrails. The attack combines simple evasion techniques—homoglyph substitution, narrative framing, and prompt decomposition—that systematically bypass safety training. This reveals that AI safety requires multi-layered defenses beyond the model itself, as guardrails remain a single point of failure against determined researchers.