Anthropic’s Claude Fable 5 Jailbroken to Generate Stack Exploits

Within days of Anthropic releasing Claude Fable 5, a public red-team jailbreak by "Pliny the Liberator" used coordinated multi-agent attacks, Unicode/homoglyph evasion, narrative framing, and decomposition/recomposition techniques to bypass safety classifiers, obtain harmful technical outputs (including exploit development steps and illicit chemistry procedures), and leak the model's ~120,000-character system prompt—highlighting risks in classifier-based routing and multi-model pipelines.